diff --git a/.project/STATUS.md b/.project/STATUS.md
index 8638c8e..d4bc060 100644
--- a/.project/STATUS.md
+++ b/.project/STATUS.md
@@ -1,5 +1,7 @@
 # artifact-driven-dev — Project Status
 
+_Updated: 2026-07-25 (**PRs #24/#26 merged + reinstall at `541b283` + first live STATUS.md prune** — both feature PRs squash-merged to `main`: #24 (`/ardd-plan --slate` spans features + open feedback, with the review-hardened deterministic bucket precedence/transitive bundles/stable ordering) and #26 (`status_history_keep` keep-last-N STATUS.md pruning, with the review-hardened fail-closed `status-prune.sh`, the shared ≤4-digit validator contract, and the new `ardd-state.sh unstamp` return-to-unbounded path); the `status-md-trim-prune-history` register flip rode #26 (`implemented`, tasks `completed` 15/15). `/ardd-update` reinstalled from this self-hosted checkout at `541b283` (0d60d69 → 541b283; no migrations pending; `status-prune.sh` + `unstamp` now installed), and its backfill ask stamped `status_history_keep: 5` — this very entry's write is the field's first dogfood: the prune below cuts the chronology to its newest 5 blocks, with the full history in git (`git log -p .project/STATUS.md`). Current state on `main`: constitution stable ✅; 0 open defects (last verified 2026-07-12); 0 open feedback; register 3 backlogged · 0 planned · 0 tasked · 42 implemented · 1 retired; no ready/in-progress tasks; no worktrees; no orphaned flips. Recommended next step: merge this update PR, then `/ardd-plan --slate` over the 3 backlogged slugs (`ci-status-refresh-on-main-coll`, `delegate-model-routing`, `plan-time-complexity-stamps`).)_
+
 _Updated: 2026-07-24 (**implemented `ardd-plan-slate-mixed-feature` on branch `feat/ardd-plan-slate-mixed`** — `/ardd-implement` executed all 10 tasks of `tasks-ardd-plan-slate-mixed-feature-4469.md` (`status: completed`, 10/10) inline on this branch: `/ardd-plan --slate` now enumerates open feedback files alongside backlogged features (each feedback *file* is one slate item), grades/relates them on the same two axes, adds the `## Reconsidered`-vs-slated-feature dependency-edge heuristic, and emits mixed `/ardd-plan <slug|feedback-*.md>` recommendations — all while staying read-only (slate never marks or flips feedback). Edits: `skills/ardd-plan/SKILL.md` (Usage `--slate` + slate steps 1–5 + next-step), `skills/ardd-feedback/SKILL.md` (Consumption note), `docs/reference/skills/ardd-plan.md` (hand-written body). No scripts or `lint-project.sh` enum changes — the scoped-run grammar already supported mixed calls. `lint-docs.sh` + `lint-project.sh` clean. Feature flipped `tasked → implemented` (rides this branch; lands when its PR merges). Sibling check: `all_complete=true`. This branch is one of two forked from `main` for the two independent features (the other, `feat/status-md-trim-prune`, carries `status-md-trim-prune-history` planning, still `tasked`). Recommended next step: push `feat/ardd-plan-slate-mixed` + open its draft PR (collaborative mode — awaiting user confirmation to push). Prior blocks preserved verbatim below.)_
 
 _Updated: 2026-07-24 (**delegated `/ardd-implement` run — docs-sweep-feedback fixes completed on this branch (PR #19)** — all 6 tasks of `tasks-chore-docs-sweep-feedback-745e.md` are complete (`status: completed`, 6/6): T001-T005 fixed the five stale `docs/reference/skills/*.md` hand-written passages (`ardd-defects` reconciliation spot-check clause, `ardd-implement` Pre-flight four-path coverage + artifacts-dir ask, `ardd-init` overclaim-guard distinction, `ardd-plan` `plan_preview_editor` coverage, `ardd-update` Claude pre-`--harness` fallback note), T006 verified `lint-docs.sh` green. All 5 bound feedback files remain `planned` against `plan-chore-docs-sweep-feedback-2026-07-24-8c58.md`; no feature slugs (`features: []`), so no register flip. State rides this branch (`worktree-agent-a9ab51aacfd6ed53a`) per worktree-native state and lands when PR #19 merges — the `ready` copy of this tasks file on `chore/docs-sweep-feedback` (PR #20) reconciles to `completed` at that point. This entry refreshes the branch's stale pre-implementation STATUS.md at the user's request; the prior blocks below are preserved verbatim.)_
@@ -8,1928 +10,3 @@ _Updated: 2026-07-24 (**`/ardd-plan` — docs-sweep feedback planned and tasked,
 
 _Updated: 2026-07-22 (**`/ardd-plan codex-second-harness-support` — closeout plan approved and tasked** — audited the stale 2026-07-15 draft plan (`plan-codex-second-harness-support-2026-07-15-f837.md`) against the current codebase before drafting anything: nearly its entire 5-phase scope was already shipped since then under other feature slugs — `install.sh --harness codex`, `new.sh --harness codex`, the `$ardd-*` invocation-sigil rewrite (feature `35f6`), `.worktreeinclude` handling (feature `multi-harness-install-metadata`), and even the old plan's own "mandatory first implementation task" (a live Codex chaining smoke test) — already run as `dev-notes/scenario-runs/2026-07-22-c0de/`. The constitution's "Multi-harness install" section already accurately documents current behavior; no correction needed. Superseded the stale draft (`draft` → `superseded`). Drafted, approved, and tasked a small closeout plan instead — `plan-codex-second-harness-support-2026-07-22-4941.md`, 2 phases/4 tasks in `tasks-codex-second-harness-support-d2fa.md` (`ready`) — covering the one real remaining gap: a `templates/AGENTS.md` pointer file + never-clobber `install.sh` wiring (mirroring the existing `.github/badges/ardd-icon.svg` idiom) for Codex installs, test-first, plus a one-line `docs/install.md` mention. Feature `codex-second-harness-support`: `backlogged` → `tasked`. No open feedback re-opened (the declined codex-scenario-sweep items stay declined, untracked by any plan, out of scope here). `lint-project` clean. **Uncommitted from the prior `/ardd-update` run**: `.project/ardd-version.md`, `.project/README.md`, `README.md`, `.github/workflows/ardd-badge.yml`, `.github/badges/` — still the user's to commit, unrelated to this plan/tasks work.) Prior update, 2026-07-22 (**`/ardd-update` — self-hosted reinstall + dynamic version badge** — this repo's `Source-Path` resolves to itself (a live dev-mode checkout, since this repo is its own ArDD source); user confirmed proceeding despite the dev-mode warning. `ardd-update-check.sh` reported `self-hosted` (tip comparison is meaningless here); the real signal was `.project/ardd-version.md` being stale (`5f47952`, 2026-07-21) against the current tree (`6895df7`, 2026-07-22, post plan-preview-editor-option merge). Reinstalled via `./install.sh .` — no pending migrations, `.project/ardd-version.md` and `.project/README.md` refreshed. Accepted the dynamic version-badge offer (README had no `ardd-badge-version-start` marker): re-ran with `ARDD_VERSION_BADGE=1`, generating `.github/workflows/ardd-badge.yml` + `.github/badges/{ardd-icon.svg,ardd-version.json}`; pasted the split badge snippet into `README.md` after the sponsor badge on confirm-with-diff (the `logo=...PLACEHOLDER` param is a known, already-documented limitation per `templates/badge-shieldcn.md`, not a new defect). All four workflow fields (`workflow_mode` absent/solo, `next_step_prompt`, `delegation`, `merge_policy`) were already set — no `--reconfigure` passed, nothing backfilled. **Not yet committed**: `.project/ardd-version.md`, `.project/README.md`, `README.md`, `.github/workflows/ardd-badge.yml`, `.github/badges/` — the user's to commit. No feature/tasks/feedback state changed by this run.) Prior update, 2026-07-22 (**`/ardd-implement` coordinator — plan-preview-editor-option implemented and merged** — the delegated `worktree-agent-a15ff0c2006ec5a70` run completed all 7 tasks and reported back; merge was a clean fast-forward (`797bd24..cf002f9`), `core.bare` unaffected, worktree reaped. Shipped: `lint-project.sh` validates the new `plan_preview_editor` frontmatter field (non-empty, must contain `{path}`) with a new bad-fixture case; `ardd-state.sh stamp` accepts `plan_preview_editor <template>` with the same free-text validation (not an enum, unlike every other workflow field); `/ardd-plan` step 10's preliminary checkpoint question now offers an "open in editor" branch alongside the existing browser-preview offer, additive and independently configurable; `docs/reference/configuration.md` and `docs/reference/scripts.md` document the new field. Full 46-script regression suite verified green (this coordinator ran every `scripts/test-*.sh` directly against the merged worktree state before merging, not just trusting the subagent's report). Feature `plan-preview-editor-option`: `tasked` → `implemented`. Open feedback: 0. Backlog now 1 backlogged (`codex-second-harness-support`) · 0 planned · 0 tasked · 39 implemented · 1 retired.) Prior update, 2026-07-22 (**`/ardd-plan plan-preview-editor-option` — approved and tasked** — targeted the sole substantive backlogged feature (`plan-preview-editor-option`; the other backlogged entry, `codex-second-harness-support`, was left alone). No artifact changes needed: the feature is a workflow field addition (`plan_preview_editor`), the same class as `plan_preview`/`delegation`/`merge_policy` that the Governance Exception already exempts from constitution body edits/SIRs — step 3's artifact-impact table found nothing to touch. Drafted, approved (no browser preview requested), and tasked `plan-plan-preview-editor-option-2026-07-22-3276.md` — 3 phases/7 tasks in `tasks-plan-preview-editor-option-8805.md` (`ready`). Design: a new optional constitution frontmatter field `plan_preview_editor` holds a shell command template containing a `{path}` placeholder (e.g. `code {path}`), validated in `lint-project.sh` and accepted by `ardd-state.sh stamp` (non-empty + contains `{path}` — a new free-text validation shape, not an enum, unlike every other existing workflow field); `/ardd-plan` step 10's preliminary browser-preview question gains an "open in editor" branch, additive to (never replacing) the existing `plan_preview` behavior. Phase 1 is schema/state-script work, test-first; Phase 2 wires the skill prose; Phase 3 updates `docs/reference/configuration.md` + `docs/reference/scripts.md`. One open feedback file (`feedback-codex-scenario-sweep-findings-2eee.md`, 3 items — ardd-update wording, ardd-defects dropped-defect bug, ardd-status over-compression) was in scope by default (no feedback-scope argument was passed) but was unrelated to this feature; per user instruction all 3 items were declined in this run rather than bundled in, and the file is now `planned` with all items marked `[-]` — **note: those 3 bugs/UX issues have no other plan tracking them now** and may need re-filing as fresh feedback if still wanted. Feature `plan-preview-editor-option`: `backlogged` → `tasked`. `lint-project` clean.) Prior update, 2026-07-22 (**`/ardd-implement` coordinator — 096b (scenario-sweep fixes) implemented and merged** — the delegated `worktree-agent-a0a55c56c6cef36a2` run completed all 6 tasks and reported back; merge was a clean fast-forward (`cabb21a..9310fa4`), `core.bare` unaffected, worktree reaped. All 5 accepted findings fixed: F001 `/ardd-update`'s pre-reinstall `--harness` guard now covers `HARNESS=claude` too, gracefully omitting the flag rather than crashing against a pre-`--harness` source; F002 `scripts/ardd-state.sh` (the real source — the subagent correctly identified `.claude/skills/ardd-scripts/ardd-state.sh` as a gitignored installed copy and edited the source instead) gained an `abspath()` helper applied to `feature-create`/`stamp`/`tasks-flip`/`task-check`/`feature-flip`/`feedback-mark`/`feedback-planned`, printing the resolved absolute path in every write confirmation; F003 `/ardd-init`'s reverse-engineer step now guards against unverified universal-coverage claims; F004 `/ardd-plan`'s task-drafting guidance now checks whether a TDD red-first task's precondition already landed earlier in the same plan; F006 `/ardd-status`'s own Feature Backlog format now states explicitly that the four core buckets always print, even at zero. Test results: `scripts/test-ardd-state.sh` 131/131 green (including 2 new red-first cases confirmed genuinely red before the fix); full 44-script regression suite green, re-verified independently by this coordinator (not just trusting the subagent's report) — behavior-preservation confirmed, only confirmation-line output changed. **One scoped deviation worth a note, not a defect**: the subagent deliberately left `mint` unchanged (writes no file — its bare-filename stdout is a data contract consumed verbatim by other skills, so adding a path would break callers) and left `plan-flip`/`feature-field` unchanged (real write-performers, but outside the plan's literal enumerated subcommand list) — flagged as an observed residual gap under F002's "every write-performing subcommand" framing, worth a small follow-up someday, not urgent. No feature flip (`features: []`). Tasks file `completed`. Open feedback: 0 (mine). Prior update, 2026-07-22 (**`/ardd-plan feedback-scenario-sweep-2026-07-21-d816-0f18.md` — approved and tasked** — drafted, approved, and tasked `plan-096b-2026-07-22-479f.md`: 2 phases/6 tasks in `tasks-096b-bfad.md` (`ready`). Fixes the 5 accepted findings from the 2026-07-21-d816 full-tier scenario sweep — F001 `/ardd-update`'s harness-compat guard only checked `HARNESS=codex`, so a plain-claude reinstall against a pre-`--harness` source still crashed passing an unrecognized flag; F002 `ardd-scripts` write-performing subcommands will start printing the resolved absolute path they wrote, turning a silent wrong-CWD write into an immediately visible one (mechanized per Principle II, not a prose reminder); F003 `/ardd-init`'s reverse-engineer step will get guidance against unverified universal-coverage claims; F004 `/ardd-plan`'s task-drafting guidance will avoid unhonorable TDD-red-first phrasing when a precondition already landed earlier in the same plan (a recurring 2026-07-15 finding class); F006 `/ardd-status`'s own Feature Backlog format gets one clarifying sentence that the four core buckets always print, even at zero. F005 (badge snippet's `PLACEHOLDER`) declined — already deliberately documented in `templates/badge-shieldcn.md`'s own header from a prior live A/B test confirming shieldcn.dev has no working substitute. Each fix was grounded against the real current file/line text before drafting, not assumed. No feature slugs (`features: []`). Feedback file now `planned`. **Notable this session: this repo is being worked concurrently by multiple live agent sessions** (2+ other Claude Code processes, one running since the prior week; a live Codex session) sharing this same checkout — a prior `STATUS.md` commit from this session raced with another session's write and lost, silently landing the other session's content under this session's own commit message. This `_Updated` entry is a fresh regeneration against current disk state (not a hand-reconciliation of the two prior versions) per the repo's own disposable-report convention — if it collides with another concurrent write, take either side, the next `/ardd-status` regenerates from disk again. A second, unrelated feedback file — `feedback-codex-scenario-sweep-findings-2eee.md` — appeared on disk mid-session from what is almost certainly that concurrent Codex session's own `/ardd-feedback` capture (references a scenario-sweep run `2026-07-22-c0de` this session never dispatched, uses Codex's `$ardd-*` syntax); left entirely untouched, not committed, not counted in the open-feedback tally below — not this session's to touch. Prior update, 2026-07-22 (**`/ardd-feedback` — Codex scenario sweep findings captured** — logged `feedback-codex-scenario-sweep-findings-2eee.md` (`open`, 2 bugs + 1 UX) from Codex sweep run `dev-notes/scenario-runs/2026-07-22-c0de`: F001 fixes `$ardd-update`'s Codex pre-harness warning wording (`.agents/skills` vs `.claude/skills`), F002 keeps `$ardd-defects` from dropping still-open known defects / under-reporting `STATUS.md`, F003 keeps `$ardd-status` from erasing useful lived-in re-entry chronology. `lint-project` clean. Open feedback now 3 files. Prior update below.)_
 
-_Updated: 2026-07-22 (**`/ardd-implement` coordinator — 35f6 (Codex `$ardd-*` skill-sigil fix) implemented and merged** — the delegated `worktree-agent-a4c94765b3a3e6564` run completed all 4 tasks and reported back; merge was a clean fast-forward (`824ac30..9dee9a3`), `core.bare` unaffected, worktree reaped. `install_skill_file()` in `install.sh` now carries a codex-only second `sed` pass rewriting genuine `/ardd-<name>` invocation references to `$ardd-<name>` — boundary check (backtick/space/`(`/`"`/start-of-line) plus an explicit 14-name alternation (never a wildcard), confirmed to leave path references, `ardd-state.sh`, and `ardd-update-check.sh` untouched, and to never over-match `/ardd-update` into `/ardd-update-check`. Red-first: `scripts/test-install-skill-sigil.sh` failed on exactly the two invocation-rewrite assertions before the fix (one self-inflicted test bug caught and fixed during the red phase — an assertion had the wrong expected path). All 8 assertions green post-fix; the Claude harness path is confirmed byte-identical to source (no behavior change for existing Claude consumers); full regression suite 46/46 green (re-verified independently by this coordinator, not just trusting the subagent's report). No feature flip (`features: []`). Tasks file `completed`. Open feedback: 1 (mine). Prior update, 2026-07-22 (**`/ardd-plan feedback-ardd-backlog-codex-syntax-1c0f.md` — approved and tasked** — drafted, approved, and tasked `plan-35f6-2026-07-22-93b3.md`: 2 phases/4 tasks in `tasks-35f6-7368.md` (`ready`). The fix scope turned out much narrower than 327 raw `/ardd-*` occurrences suggested: `install.sh`'s `install_skill_file()` already has exactly this class of mechanism (a harness-conditional `sed` rewrite for `.claude/skills` path prose, with a comment explicitly noting invocation syntax was deliberately left alone) — this plan extends that same function with a codex-only second pass, rather than hand-editing 14 prose files. Design grounded empirically this session: grepped the real pattern space (327 genuine invocation occurrences vs. 88 path occurrences across all 14 skills), identified five non-invocation `ardd-*` names that must never be touched (`ardd-state`, `ardd-update-check`, `ardd-version`, `ardd-scripts`, `ardd-artifact-templates` — confirmed none currently appear in an invocation-looking context), and — prompted by a live user question — separately verified the fix's boundary+allowlist regex structurally cannot touch any built-in harness command (`/exit`, `/init`, etc.) since those don't contain the literal `ardd-` substring after the slash. Test-first: T001/T002 write and red-check a new `scripts/test-install-skill-sigil.sh` before T003/T004 implement the codex-only `sed` pass and go green. No feature slugs (`features: []`). Feedback file now `planned`. Open feedback now 1 file (mine; a second, `feedback-codex-scenario-sweep-findings-2eee.md`, remains the untouched third-party file noted below). Prior update, 2026-07-22 (**b860 Codex launch-tty fix — live-verified** — the user ran `codex '$ardd-status'` in a real terminal against the fixed `new.sh`; Codex launched correctly and received `$ardd-status` as its first turn. This confirms the `<> /dev/tty` redirect is the correct fix and resolves the plan's flagged Open Question — no session (this coordinator or the delegated worktree subagent) had a controlling tty to confirm this themselves, so this was genuinely unverified until now. The b860 fix (merged `587d794`) is fully confirmed working, not just statically tested. Prior update, 2026-07-22 (**`/ardd-implement` coordinator — b860 (Codex launch-tty fix) implemented and merged** — the delegated `worktree-agent-a2f2806ef246ff447` run completed both tasks and reported back; merge was a clean fast-forward (`e923fbd..587d794`), `core.bare` unaffected, worktree reaped. `new.sh`'s `launch()` now branches its `exec` on `$harness`: Claude Code's path is unchanged (still no stdin redirect); Codex's path adds `<> /dev/tty` (read-write), since Codex has no `isTTY`-checked `/dev/tty` fallback of its own and errors immediately without a real terminal on stdin — confirmed empirically this session (`echo "" | codex '...' </dev/null` reproduces `Error: stdin is not a terminal`). `test-new.sh` case 14 now asserts both branches separately (same static/no-live-tty discipline as before); full suite green, 80/80 checks. **Still open, per the plan's own flagged Open Question: this fix has NOT been live-verified in a real terminal** — neither this coordinator session nor the delegated worktree subagent has a controlling tty (`test -t 0` false in both), so only the failure could be reproduced here, never the fix. Whoever has a real terminal should run `codex '$ardd-status'` (or similar) against a real Codex-harness project before treating this as fully confirmed — Codex CLI (`codex-cli 0.144.6`) is on this machine at `~/.local/bin/codex`. No feature flip (`features: []`). Tasks file `completed`. Open feedback: 2. **Noted, not touched:** a third-party feedback file, `feedback-codex-scenario-sweep-findings-2eee.md`, appeared on disk (untracked) during this session, referencing a scenario-sweep run (`2026-07-22-c0de`) this session never dispatched and using Codex's `$ardd-*` invocation syntax — almost certainly a concurrent Codex session's own `/ardd-feedback` capture landing in this same checkout. Left entirely alone (not committed, not read further, not counted in the open-feedback tally above) — not this session's to touch. Prior update, 2026-07-22 (**`/ardd-plan feedback-new-sh-codex-launch-tty-b951.md` — approved and tasked** — drafted, approved, and tasked `plan-b860-2026-07-22-e518.md`: 1 phase/2 tasks in `tasks-b860-596c.md` (`ready`). Empirically confirmed locally (`codex-cli 0.144.6`) that piping EOF'd stdin to `codex` reproduces the exact reported error, then designed the fix: branch `launch()`'s `exec` on `$harness` — Claude Code's exec stays unredirected (unchanged, case-14-guarded), Codex's branch adds `<> /dev/tty` (the read-write redirect Codex needs since it has no `isTTY`-checked `/dev/tty` fallback of its own, unlike Claude Code). Flagged an Open Question the implementing task must resolve live: this sandbox has no controlling terminal at all, so only the *failure* could be reproduced here, not the *fix* — T001 requires empirical verification in a real terminal before the fix is considered confirmed, not just a passing static test. T002 splits `test-new.sh` case 14's guard by harness branch (same static/no-live-tty discipline the existing case already uses). No feature slugs (`features: []`). Feedback file now `planned`. Open feedback now 2 files. Prior update, 2026-07-22 (**`/ardd-feedback` — systemic Codex-syntax drift in skill prose** — reported live from a real Codex session: `/ardd-backlog` printed "When ready to design it: `/ardd-plan codex-ai-provider-support.`" — Claude's `/` sigil in a Codex session, where the correct invocation is `$ardd-plan`. Confirmed this is systemic rather than an `ardd-backlog`-only gap: every skill's terminal-handoff/cross-reference prose hardcodes the `/ardd-*` form (`skills/ardd-backlog/SKILL.md:161` is the reported instance; `ardd-plan`/`ardd-status` each contain dozens more) — only the root-side acquisition scripts (`new.sh`/`install.sh`) are harness-aware about the `/` vs `$` prefix today, the installed skill prose itself never adjusts. Logged `feedback-ardd-backlog-codex-syntax-1c0f.md` (`open`, 1 bug, flagged as needing a systemic fix — a harness-neutral phrasing convention across every skill, not a one-line patch — and worth sequencing alongside `codex-second-harness-support`'s remaining design work and `multi-harness-install-metadata`, same dual-harness-prose-drift risk class). Open feedback now 3 files. Prior update, 2026-07-22 (**`/ardd-feedback` — live Codex-harness `new.sh` bug from a real consumer install** — the user installed `--harness codex` support into a real project (assisted-review) via `curl | sh` and hit `Error: stdin is not a terminal` on the kickoff launch step: `new.sh`'s `launch()` function (`new.sh:346-357`) `exec`s the handoff tool without redirecting stdin, a design deliberately tailored to Claude Code's own `isTTY`-checked `/dev/tty` fallback behavior (per the function's own comment) — but the identical `exec` pattern is reused unconditionally for Codex, whose CLI apparently lacks that same fallback and just requires a real terminal on stdin, which a `curl | sh` pipe never provides. Logged `feedback-new-sh-codex-launch-tty-b951.md` (`open`, 1 bug) — needs a Codex-specific stdin-redirect fix or documented no-launch fallback, verified against Codex's actual stdin behavior rather than assumed from the Claude Code case. Open feedback now 2 files. Prior update, 2026-07-22 (**full-tier `/scenario-sweep S1-S8` (run `2026-07-21-d816`) + stable v1.1.0 shipped** — a full pre-release sweep across all 8 scenarios came back clean: zero crashes/hard failures, every skill ran to a coherent end or refused correctly. 6 accepted findings logged as `feedback-scenario-sweep-2026-07-21-d816-0f18.md` (`open`, 3 bugs + 3 UX — none release-blocking): F001 `/ardd-update` unconditionally passes `--harness <harness>` to `install.sh` but its compat guard only checks `HARNESS=codex`, hard-failing a reinstall against a pre-`--harness` stable tag (S3); F002 `ardd-scripts` CWD-relative subcommands resolve `.project/` against the invoking shell's CWD even when invoked by absolute path — silent wrong-target write (S6); F003 a reverse-engineered constitution overclaimed a principle's coverage, self-caught by a later `/ardd-defects` run rather than persisting silently (S2); F004 a drafted task carried unhonorable TDD-red-first phrasing, recurrence of a 2026-07-15 finding class (S5); F005 the badge snippet's `logo=` param carries a literal `PLACEHOLDER` string (S1); F006 `/ardd-status`'s Feature Backlog line omits explicit "0 planned" (S7). Mid-sweep, S4's dispatched subagent got stuck in a resume loop — completed real collaborative-mode implementation work (branch gate, real code, task/register flips, all correct) but then re-inspected already-correct sandbox state across 3+ resumes without writing report progress; killed the stuck background tasks, root-caused via a fable-agent consult (resume-state amnesia — no rule treated the report file as a position marker), and hardened `tests/scenarios/GUARDRAILS.md` (new rules 12-13: report file is the resume position-marker, max-2-inspection-commands bail rule) plus `.claude/skills/scenario-sweep/SKILL.md`'s dispatch closing reminder — committed and pushed (`85407e4`) ahead of the sweep's own fix loop, since it's harness tooling, not a consumer-facing finding. S4 finished by hand against the sandbox: all 6 checklist items passed, including confirming the `gh pr create` failure-handling fix from 2026-07-15 now works correctly. Full cost/triage record: `dev-notes/scenario-runs/2026-07-21-d816/{RUN,TRIAGE}.md`. With triage clean, pushed `main` (beta `v1.1.0-beta.x` published) and dispatched `stable-release.yml` (bump: minor, since v1.0.4 shipped multi-harness install support as a new capability, not just fixes) — **v1.1.0 is now Latest**, carrying multi-harness install, badge/shieldcn styling, brand refresh, and the accumulated post-v1.0.4 fixes to stable consumers. The 6 sweep findings above remain open for the next `/ardd-plan` to consume. Prior update, 2026-07-21 (**`/ardd-implement` coordinator — new-sh-codex-docs implemented and merged** — the delegated `worktree-agent-a96704d0e644f9288` run completed its single task and reported back; merge was a clean fast-forward this time (`cc522a9..4f0177c`, no register-flip race like the earlier changelog-precommit merge — this run had no bound features to flip). `docs/install.md` now documents `new.sh --harness claude|codex` (the flag, the `ask_harness` interactive prompt with its `/dev/tty` fallback discipline, the pass-through to `install.sh --harness`) right alongside the existing `--kickoff`/`--no-kickoff` bullet, cross-referencing the existing "Codex CLI: `install.sh --harness codex`" section. `lint-docs`/`lint-project` both clean. Worktree reaped. Open feedback: 0. Ready tasks files: 0. Prior update, same day, **`/ardd-plan feedback-new-sh-codex-undocumented-b8f0.md` + a further docs-sweep generalization** — approved and tasked `plan-new-sh-codex-docs-2026-07-21-302a.md`: 1 task/1 phase in `tasks-new-sh-codex-docs-2770.md` (`ready`) — add `new.sh --harness claude|codex` (the flag, the interactive `ask_harness` prompt, the pass-through to `install.sh --harness`) to `docs/install.md`'s existing `new.sh` flag documentation, cross-referencing the existing "Codex CLI: `install.sh --harness codex`" section. No bound features (`features: []`). Feedback file now `planned`. Meanwhile, mid-run, the user asked for a further `docs-sweep` generalization: rather than hardcoding `new.sh`/`install.sh` as the only "scripts in scope," dispatched a second subagent to add a discovered-not-hardcoded standalone-script check — a new §2c grep-based procedure (skill-invoked set from `skills/*/SKILL.md`, minus CI-workflow-invoked and settings.json-hook-wired scripts, minus `test-*`/`lint-*`) that flags any newly-added standalone script with zero mentions anywhere in `docs/install.md`/`README.md`/`USAGE.md`/`docs/concepts.md` — a coarser "does anything reference this script's existence at all" net, distinct from the deep per-script audit `new.sh`/`install.sh` already get. Tested against this repo before committing: the classification correctly excludes all 20 skill-invoked scripts, `next-version.sh`/`release-notes.sh`/`smoke-assert.sh` (CI-invoked), and `hook-lint-on-write.sh` (settings.json hook-wired) — leaving only `gen-skill-docs.sh` as a genuine standing candidate (not hand-waved away; it really isn't documented or CI-wired). Committed `937479e`, signed. Prior update, same day, **pre-commit hook fix + docs-sweep scope extension + new.sh docs gap** — three follow-ups from a live post-implementation check of this session's `pre-commit-hook-scoping-expans` work. First, the user flagged the scoping as "overly cautious": rewrote `hooks/pre-commit` directly (no plan/tasks ceremony — a source-side dev-tooling fix) so paired `test-X.sh`/`scripts/X.sh` checks run ONLY when that specific pair changes (dropped the old "unknown subject, run it anyway" fallback that was silently forcing every migration/install-family test to run on any unrelated commit), and any staged path outside every mapped area now runs a small curated regression suite (`lint-docs.sh`, `lint-project.sh`, `lint-templates-yaml.sh`, `test-hooks-pre-commit.sh`) instead of the entire ~45-script suite — `ARDD_HOOK_ALL=1`/empty-staged-list still force the literal full run. 6 new fixture cases added to `test-hooks-pre-commit.sh`; measured effect: an unmapped-content commit dropped from running the whole suite (~50s+) to ~19s. Committed `480680d`, pushed. Second, verified `new.sh` supports `--harness codex` (interactive prompt + flag, passed through to `install.sh --harness`) but found `docs/install.md` never mentions this — filed `feedback-new-sh-codex-undocumented-b8f0.md` (1 bug, open). Third, investigated why `/docs-sweep` never caught that gap: its entire scope-resolution and per-item procedure is anchored to `skills/*/SKILL.md` — `new.sh`/`install.sh` are root-level scripts, not skills, so nothing in its procedure ever looked at them. Dispatched a subagent to extend `docs-sweep`'s scope: a parallel "scripts in scope" set (`new.sh`, `install.sh` only — not all of `scripts/`) driven by the same last-stable-tag diff and `--all` rules, a new §2b per-script judgment procedure (ground truth = the script's own source/header/usage text, doc target = `docs/install.md` primarily), and a widened triage table header. Committed `9b1c5fd`, signed, verified. Both fixes pushed to `main`; `be71d6c..0a4df4f`. Prior update, same day, **`/ardd-implement` coordinator — changelog-precommit implemented and merged** — the delegated `worktree-agent-af5346a43bebe9e32` run completed all 13 tasks of `tasks-changelog-precommit-0966.md` and reported back; merge hit a real conflict (both feature files' `status:` line — main said `tasked`, the branch said `implemented`), so per `merge_policy: auto` the coordinator aborted and asked rather than auto-resolving; user confirmed taking the branch's `implemented` value (`git merge -X theirs`, then committed with the on-disk signing key after 1Password's agent errored locked). Merged `ff3dabb`; full suite green; worktree reaped. Shipped: new source-side `scripts/release-notes.sh` (+ hermetic `scripts/test-release-notes.sh`) regenerating `docs/release-notes.md` from `gh api` releases, wired into `stable-release.yml` right after a release is cut and into a new `release-notes` CI lint job; `hooks/pre-commit`'s staged-path scoping extended — `lint-templates-yaml.sh` now runs on `.github/workflows/`/`templates/` changes, `tests/fixtures/` maps to its two consumers, `CLAUDE.md`/`dev-notes/*`/`tests/scenarios/*`/`mkdocs.yml`/`.gitignore`/`.worktreeinclude` are recognized no-check paths — with matching fixture cases in `scripts/test-hooks-pre-commit.sh`; live A/B-tested shieldcn.dev's `dynamic/json` logo param and confirmed it does **not** accept a base64 data URI (silently ignored; a named slug renders differently) — `templates/badge-shieldcn.md`'s header/caveat updated with this confirmed finding, `PLACEHOLDER` deliberately left in place since no working substitute exists; this repo's own `.worktreeinclude` now covers `.agents/skills/scenario-sweep/`; `/ardd-implement`'s post-delegation coordinator check (`skills/ardd-implement/SKILL.md` step 3) now also restores `core.hooksPath` alongside the existing `core.bare` check, mentioned in `CLAUDE.md` and decision record 0001. Blocker the subagent hit and fixed: the register was left mid-sequence (plan `approved`, both features still `backlogged`) — a gap from the coordinator's earlier auto-commit-before-delegating step only covering the plan/tasks files, not the intervening `ardd-state.sh` register writes; repaired in-worktree before proceeding (and again surfaced as the merge conflict above, from main's own uncommitted-then-committed copy of the same fields diverging). Both features `tasked` → `implemented`. Open feedback: 0. Ready tasks files: 0. Next `/ardd-status` run reflects a fully quiescent backlog (2 backlogged, 0 tasked, 38 implemented, 1 retired). Prior update, same day, **`/ardd-plan` — changelog-from-github-releases + pre-commit-hook-scoping-expans planned and tasked** — a bare `/ardd-plan` offered the step-1a picker over the then-4-item backlog and the 2 open feedback files; user selected both `changelog-from-github-releases` and `pre-commit-hook-scoping-expans` plus both feedback files. No artifact changes needed (only `constitution.md` exists; neither feature touches a declared principle). Drafted, approved, and tasked `plan-changelog-precommit-2026-07-21-b716.md` — 13 tasks/3 phases in `tasks-changelog-precommit-0966.md` (`ready`): Phase 1 a new source-side `scripts/release-notes.sh` (+ fixture test + CI job) regenerating `docs/release-notes.md` from `gh release list`/`gh api`, wired into `stable-release.yml` right after a release is cut; Phase 2 extends `hooks/pre-commit`'s staged-path scoping — `lint-templates-yaml.sh` added to the check loop (`.github/workflows/`+`templates/` subjects), `tests/fixtures/` mapped to its two consumers, `CLAUDE.md`/`dev-notes/*`/`tests/scenarios/*`/`mkdocs.yml`/`.gitignore`/`.worktreeinclude` recognized as no-check paths, plus new `test-hooks-pre-commit.sh` fixture cases pinning each mapping; Phase 3 folds in the two open feedback files' 3 items as independent follow-up tasks — the shieldcn badge logo `PLACEHOLDER` (verify shieldcn.dev's `dynamic/json` logo param, then replace with the real base64-encoded `templates/ardd-icon.svg` value), this repo's own `.worktreeinclude` gap for `.agents/skills/scenario-sweep/` (source-side dev tooling, not an `install.sh` target-facing change), and `/ardd-implement`'s missing `core.hooksPath` restoration alongside its existing `core.bare` check (+ doc mentions in CLAUDE.md and decision record 0001). Both features `backlogged` → `tasked`; both feedback files now `planned` — open feedback is 0. Ran `/ardd-status` per the terminal handoff. Prior update, same day, **badge-style-variant-option planned and tasked** — amended the badge brand colour to `#2F4858` (was `#7C3AED`) across `docs/reference/configuration.md`, `templates/ardd-badge.json`, `templates/ardd-badge-workflow.yml`, and `scripts/test-install-version-badge.sh`'s case16 assertion; regression suite green. Then a bare `/ardd-plan` picked `badge-style-variant-option` from the 3-item backlog: drafted, approved, and tasked `plan-badge-style-variant-option-2026-07-21-982c.md` — 10 tasks/5 phases (incl. a standalone Phase 0) in `tasks-badge-style-variant-option-7230.md` (`ready`). Design: a new `EXISTING_SHIELDS_IO` README detection (marker-excluded grep for `img.shields.io`) alongside the existing `BADGE_FAMILY` marker detection — shieldcn.dev becomes the default install-time badge offer (mirroring this repo's own sponsor badge, `README.md:13`), shields.io the fallback when the target README already carries non-ArDD shields.io badges; new `templates/badge-shieldcn.md` (3 shapes) ships alongside `templates/badge.md`; open question flagged on shieldcn.dev's exact `dynamic/json` query syntax (inferred from the shields.io analog, not yet independently verified). Mid-run catch: step 4's open-feedback check had missed `feedback-badge-style-clash-7db2.md` (a strict-equality status match didn't tolerate its trailing frontmatter comment) — surfaced it after tasking, confirmed with the user, marked F001 incorporated, bound to this plan (now `planned`), and added a standalone Phase 0/T010 for its interim-caveat ask (shields.io-vs-shieldcn clash note in the *current* offer text, landing independent of and before the rest of the plan). Feature `badge-style-variant-option`: `backlogged` → `tasked`. Open feedback now 1 file (`feedback-hookspath-side-effect-c707.md`). Prior update, same day, **multi-harness plan implemented and merged** — the delegated worktree run completed all 12 tasks of `tasks-multi-harness-49bb.md` (full suite green), merged to `main` as d41b193, worktree reaped; feature `multi-harness-install-metadata` is **implemented**. Shipped: `Harnesses:` union line in ardd-version.md (order-normalized, absent = claude; dual installs both orders preserve the sibling), dev-mode reinstall records `Channel: dev` and drops Source-Ref, per-harness bounded `.worktreeinclude`/gitignore guidance, reviewer guide lists every installed root, **pre-commit hook staged-path scoping** (P90: `.project/`-only commits ~117s → ~8s; `ARDD_HOOK_ALL=1` override), ardd-update reference/docs/USAGE `--harness codex` documentation, GUARDRAILS cwd-safety rules 7–8, ardd-plan two-sequential-prompts clause, S2 cold-fixture cleanse, S7 reviewer-guide check, init/backlog collaborative scaffold notes. Subagent also fixed a real fixture-isolation bug (`test-hooks-pre-commit.sh` missing the GIT_DIR unset guard). Same session: **brand refresh** (70f9c2a) — favicon + badge mark redesigned (orbit ring + 4 triangles around a small filled rect; heavier legibility variant; badge icon in dark-context whites); note existing consumers keep their never-clobbered `.github/badges/ardd-icon.svg` until refreshed. Second inbox drain (8a666ed): feedback-badge-style-clash-7db2 (open) + backlogged `badge-style-variant-option`, then revised by user decision (fcd287f): **shieldcn becomes the default badge offer**, shields.io the detection-based fallback; both templates ship; agents may adapt to detected style; other-style users invited to contribute templates upstream. Prior update, same day: **/ardd-plan multi-harness — approved and tasked** — constitution bumped to v1.13.0 (new Multi-harness standing decision: dual Claude+Codex installs are first-class; shared `.project/` install metadata represents the full harness set, preserve-on-reinstall, never last-writer-wins); `plan-multi-harness-2026-07-21-76ba.md` approved, `tasks-multi-harness-49bb.md` **ready** (12 tasks / 4 phases: harness metadata in install.sh, pre-commit staged-path scoping for P90 [~117s → ~7–10s common case, per the 2026-07-21 timing investigation], docs drift, guardrails/scenario/prose hardening); feature `multi-harness-install-metadata` (inbox-drained this session) flipped `tasked`. Consumed ALL 5 open feedback files (12 items, all accepted): docs-sweep-28d0, sweep-cdf0-878c, guardrails-5f18, plan-preview-prompts-a91e, pre-commit-hook-p90-1e7b — open feedback is now 0. Committed 91a2a40 (plus 6527695 backlog entry, 3244f70 status, b1c06bc docs-sweep feedback). Prior update, same day: **codex foundation merged to main** — the user merged `codex/port-foundation` (8a302c8: `install.sh --harness codex` support); the worktree at `…/artifact-driven-dev-codex-port` is now merged + clean → **reapable** (`worktree-reap.sh` candidate; reap pending). Register state deliberate: `codex-second-harness-support` stays `backlogged` and its plan `draft` — the merge is the foundation slice, not the full feature. Same session, `/docs-sweep` (scope: v1.0.3..HEAD, sole changed skill `ardd-update`) logged `feedback-docs-sweep-ardd-update-harness-28d0.md` (open, 3 bugs): the ardd-update reference page's step-4 body lacks the new harness-preservation behavior, still claims "your README is never edited" against the confirm-with-diff posture, and `install.sh --harness codex` is undocumented in all human-facing docs. Open feedback now 6 files. Prior update, same day, **v1.0.4 shipped** — pushed main (02c9692..68970e8, 13 commits; beta-release published v1.0.4-beta.1, lint green incl. the template-YAML jobs), then dispatched stable-release.yml (bump: patch): **v1.0.4 is Latest**, carrying the post-1.0.3 badge hardening to stable consumers — marker-family guards, real-default-branch workflow trigger, ssh-alias coordinate fill, confirm-with-diff README posture, amend-path prose. Every S9 gate in the arc ran clean before shipping. Prior update, same day, /ardd-feedback — logged feedback-guardrails-cwd-safety-5f18.md (open, 1 bug): the primary checkout lost its git origin remote during the S9 runs (restored by hand); likely a scenario subagent running the guardrails rule-2 cleanup outside $SCRATCH. Fix direction: GUARDRAILS.md gains a cwd-inside-$SCRATCH verification before any prescribed git mutation, structural -C/absolute-path forms, and an incident-reporting rule. Prior update, same day, /ardd-implement coordinator — the delegated badge-consumer-fixes worktree run completed all 4 tasks and merged fast-forward (5f47952); worktree reaped; dogfood refreshed. Shipped: install.sh parses any scp-style <token>:<path> remote (ssh-config aliases like github-ardd: now fill real coordinates; red-first case 22, badge suite 48/48); confirm-with-diff posture at templates/badge.md + install output + ardd-update relay (agent OFFERS the README edit with a diff, never refuses-until-overridden; script stays suggestion-only); S9 case-2 alias-remote variant. Badge file-set moved, so S9 run 2026-07-21-b788 ran against 5f47952: CLEAN PASS — ea66 F001 (alias coordinates fill) and F002 (confirm-with-diff posture on all three surfaces) verified fixed, all prior badge fixes hold; sole new finding taste-deferred (third of the bare-wrong-badge output-pairing family). Prior update, same day, /ardd-plan + /ardd-feedback — consumed feedback ea66 (both items accepted, planned): approved and tasked plan-badge-consumer-fixes-2026-07-21-1c50.md — 4 tasks/2 phases in tasks-badge-consumer-fixes-56c0.md (ready): red-first ssh-alias remote parse (scp-style <token>:<path> regardless of host token), confirm-with-diff README posture at the three badge prose sites, S9 alias-remote brief variant. Also logged feedback-plan-preview-prompts-a91e.md (open, 1 UX: preview and approval must be two sequential prompts, preview shown first — observed batched live) and backlogged plan-preview-editor-option (open-in-editor checkpoint option + configurable editor command). Prior update, same day, `/ardd-feedback` — logged
-`feedback-badge-consumer-feedback-ea66.md` (`open`, 1 bug + 1
-Reconsidered) from the first real consumer update to v1.0.3
-(yet-another-rank-games): F001 the badge coordinate autodetect misses
-scp-style SSH host-alias remotes (`github-ardd:moui72/...`) and falls
-back to placeholders — parse the path part regardless of host token;
-F002 the never-edit-README posture reads as refusal-until-overridden
-to consuming agents — revise the badge-surface prose so the agent
-offers a confirm-with-diff gate instead (install.sh itself stays
-suggestion-only). The next `/ardd-plan` picks it up. Prior update,
-2026-07-20, **v1.0.3 shipped** — with the S9 gate clean, the
-user committed the codex plan refresh (plan file only; the
-feature-entry edit was deliberately withheld), pushed `main`
-(`1803626..0fc43f6`; beta-release published `v1.0.3-beta.2`), and
-dispatched `stable-release.yml` (`bump: patch`) — **`v1.0.3` is now the
-Latest stable release**, carrying the whole badge arc to stable
-consumers: valid workflow YAML, real-default-branch trigger filter,
-marker-family guards, brand `labelColor` + `logoSvg` icon, split-badge
-default, reviewer guide, `next_step_prompt: auto`, portable
-Source-Path, and the amend-path policy. Prior update, same day, **S9
-gate run `2026-07-20-270a`: CLEAN PASS —
-the stable gate is cleared.** 8110-F001 verified fixed (the written
-badge workflow carried `branches: [master]` — the repo's real default
-branch — in every badged case, with the install output printing it);
-b8b6 F001–F003 spot-checks all hold; all 8 checklist items green;
-idempotent reruns byte-identical; wrong-badge advisory fired in all
-four combos; `/ardd-update` left badge state byte-identical. Sole
-finding: one subjective ux sequencing nit (bare wrong badge gets
-advisory + paste block), taste-deferred as a duplicate of run-7b1a's
-already-deferred note. The badge surface is release-ready: push `main`
-(publishes the next beta) and dispatch `stable-release.yml` (patch →
-v1.0.3) — both remain the user's acts. Prior update, same day, S9
-rerun + branch-filter fix — regression run
-`2026-07-20-7b1a` verified all three b8b6 guard fixes live (8/8
-checklist) but surfaced one new consumer-facing bug, triaged with the
-user: `templates/ardd-badge-workflow.yml` hardcoded `on.push.branches:
-[main]`, so master-default repos never fire the badge sync (accepted as
-`feedback-badge-workflow-branch-filter-8110.md`; two ux notes
-taste-deferred; `Channel: stable`-under-live-checkout declined as
-by-design after verifying install.sh's channel-precedence comment).
-Full loop run: plan `plan-badge-workflow-branch-2026-07-20-abf3.md`
-(approved) → delegated worktree run (red-first case 21, then install.sh
-computes `BADGE_BRANCH` and sed-substitutes it into the workflow's
-`branches:` filter on the fresh-write path; suite 45/45) → merged
-fast-forward (`f4433ed`), reaped, dogfood refreshed. **Stable gate
-still holding (user instruction)**: S9 gate run `2026-07-20-270a`
-dispatched against `f4433ed` — only a clean pass unlocks the stable
-dispatch, which remains the user's act. Also: user's
-`codex/port-foundation` worktree remains in flight (reap refuses,
-dirty); `.agents/skills/scenario-sweep/SKILL.md` gitignored on request.
-Prior update, same day, `/ardd-implement` coordinator — the delegated
-`badge-guards` worktree run completed both tasks and merged
-fast-forward (`b95dd2e`); worktree reaped (the reap correctly refused
-the user's dirty `codex/port-foundation` worktree — in-flight work,
-untouched); dogfooded install refreshed. Shipped: red-first cases 18–20
-in `test-install-version-badge.sh`, then install.sh's badge section
-gained `BADGE_FAMILY` detection (word-exact bounded regexes,
-longest-first) — env-unset installs now acknowledge any marker family
-instead of re-suggesting the static badge (b8b6-F001), pair/static
-repos get a switch-shapes note instead of the paste block (b8b6-F002),
-and the no-README pointer is flag-aware (b8b6-F003). Suite 44/44 +
-lint-docs green. Tasks file `completed`; no feature flips. **Stable
-gate (user instruction): do NOT dispatch stable-release.yml until a
-clean S9 rerun** — regression run `2026-07-20-7b1a` dispatched against
-`b95dd2e`, re-verifying F001–F003. Prior update, same day,
-`/ardd-plan feedback-s9-badge-guards-b8b6.md` —
-consumed the S9 findings (all 3 accepted, file now `planned`): drafted,
-approved, and tasked `plan-badge-guards-2026-07-20-8b60.md` — 2
-tasks/2 phases in `tasks-badge-guards-cddd.md` (`ready`, red-first).
-T001 adds three red cases to `test-install-version-badge.sh`
-(version-badged env-unset → no static re-suggestion; pair-badged
-env-set → acknowledgment instead of the paste block; flag-aware
-no-README pointer); T002 adds marker-family detection and branches both
-guards on it (word-exact marker matching — substring overlap warned).
-Post-merge regression gate: rerun `/scenario-sweep S9` before the next
-stable dispatch. No feature slugs (`features: []`). Committed to `main`
-(a6c0d73). Noted but left uncommitted: an out-of-session refresh of
-`plan-codex-second-harness-support-2026-07-15-f837.md` + its feature
-entry (a 2026-07-20 capability-baseline rewrite) sits in the working
-tree — not this session's work, folded in when its author is ready.
-Prior update, same day, badge scenario + first S9 run + `/ardd-feedback`
-— a delegated design run authored `tests/scenarios/S9.md` (badge
-matrix: 6 repo states — no-README, no-badge, current-format re-install,
-pair/static previous formats, homespun wrong badge, `/ardd-update`
-over badged — with a reusable well-formedness pass) and a new
-"change-triggered (targeted-only)" tier category mapping the badge
-file-set → `/scenario-sweep S9` (merged `6877f80`, reaped). Then S9's
-first live run (`2026-07-20-4f1a`, ~64K tokens, ~$0.20–0.80, 3m45s):
-**PASS, no hard failures** — README never edited across 11 installs,
-advisory fired in all 4 misdirected-badge runs, all well-formedness
-checks green. 3 accepted ux findings logged as
-`feedback-s9-badge-guards-b8b6.md` (`open`): F001 the static-suggestion
-guard misses `ardd-badge-version-start` so env-unset installs (incl.
-every `/ardd-update`) re-suggest the static badge to a
-correctly-badged repo; F002 the version-snippet guard misses
-`ardd-badge-pair-start` so pair-badged repos get the full paste block;
-F003 the no-README pointer ignores that `ARDD_VERSION_BADGE=1` is
-already set. Common root: both guards should treat all three marker
-families as "already badged". Earlier: pushed `main`
-(`33ac9ae..1803626`), lint green incl. the two new template-YAML CI
-jobs, **v1.0.3-beta.1 published** (stable v1.0.2 still ships the
-broken badge workflow). Prior update, same day, `/ardd-implement`
-coordinator — the delegated
-`badge-slate` worktree run completed all 7 tasks and merged
-fast-forward (`5adffc3`); worktree reaped; dogfooded install refreshed.
-Shipped: new `scripts/lint-templates-yaml.sh` + fixture test + two CI
-jobs (red-first: the lint failed on the broken template before the
-fix; self-skips locally without PyYAML — CI installs it);
-`templates/ardd-badge-workflow.yml` fixed (heredoc indented into the
-`run: |` scalar, `branches: [main]` + `workflow_dispatch`,
-`[skip ci]`) and verified by executing the extracted run script;
-`templates/ardd-icon.svg` (323-byte single-colour delta monogram);
-brand hex `#7C3AED` documented in `configuration.md` and emitted as
-`labelColor`, icon inlined as `logoSvg` via `jq --rawfile` (install.sh
-ships the icon never-clobber and seeds both fields jq-free via an awk
-JSON-escaper); `badge.md` restructured to three shapes with the split
-badge as recommended default + the bf75 renderer caveat and base64
-data-URI recipe; install.sh's printed snippet now the split badge with
-all four existing guard behaviors preserved. Full suite green on every
-commit. All three badge features `tasked → implemented`. Prior update,
-same day, `/ardd-plan` badge slate — consumed both open
-badge feedback files (`ff0c` 1 bug + 3 UX, `bf75` 1 UX — all 5 items
-accepted, both now `planned`) and targeted the three badge features:
-drafted, approved, and tasked `plan-badge-slate-2026-07-20-e059.md` —
-7 tasks/3 phases in `tasks-badge-slate-658d.md` (`ready`). Phase 1
-(test-first): new `lint-templates-yaml.sh` (YAML-parse over shipped
-templates + workflows, red on the currently-broken badge template) +
-the verified heredoc-indent fix, `branches: [main]` +
-`workflow_dispatch`, `[skip ci]`. Phase 2: brand hex (proposed
-`#7C3AED`) published + emitted as `labelColor`; `templates/
-ardd-icon.svg` created and inlined as `logoSvg` (file-read at
-generation, never escaped-SVG-in-heredoc); badge install test
-extended. Phase 3: `badge.md` split-badge variant (recommended
-default), pair kept, renderer endpoint-vs-dynamic-JSON caveat;
-install.sh snippet + docs. Open questions carried: final brand hex;
-split-as-printed-default. All three features `backlogged→tasked`.
-Committed to `main` (59cc519). Prior update, same day,
-`/ardd-implement` coordinator — the delegated
-`amend-path-policy` worktree run completed all 5 tasks and merged
-(`ort`, no conflicts — merge commit `05d3978`); worktree reaped;
-dogfooded install refreshed. Shipped: the **factual-corrections
-exemption** is now canonical in `templates/dot-project-readme.md`
-("Correcting a skill-written file" — body-prose fixes to
-citations/paths/symbols/typos hand-editable; frontmatter/lifecycle
-state and decisions never), echoed in ardd-feedback/ardd-plan prose,
-`docs/reference/project-files.md`, `docs/concepts.md` (the README
-section the plan named had moved there — recorded in the T003 commit),
-and CLAUDE.md; plus the citation-rot sweep — ardd-defects and
-ardd-feedback prose now prefer path + symbol citations, line numbers as
-supplementary hints only (ardd-lint's file:line output untouched:
-script-emitted). Tasks file `completed`; no feature flips
-(`features: []`). Both lints green; this repo's own `.project/README.md`
-regenerated with the new subsection. Prior update, same day,
-`/inbox-drain` + `/ardd-feedback` — drained one
-inbox item (three badge proposals from adopting the guidance in
-moui72/assisted-review PRs #107/#108): re-filed 3 new-capability items
-to the register — `badge-split-variant` (a single split badge as a
-third `badge.md` shape, maybe the default), `badge-brand-color-in-json`
-(canonical ArDD brand hex published and emitted as `labelColor` in the
-generated badge JSON — central brand propagation), `badge-icon-logosvg`
-(ship `templates/ardd-icon.svg`, workflow inlines it as `logoSvg` +
-base64 URL form) — all `backlogged`, cross-linked; wrote
-`feedback-badge-renderer-caveat-bf75.md` (`open`, 1 UX): `badge.md`
-implies the JSON drives appearance, untrue for dynamic-JSON renderers
-(shieldcn) — document the endpoint-vs-dynamic distinction. Inbox file
-deleted; committed. Meanwhile the delegated `amend-path-policy`
-worktree run is in flight at 1/5. Prior update, same day,
-`/ardd-plan` — consumed
-`feedback-amend-path-for-skill-written-files-2187.md` (both items
-accepted, file now `planned`): drafted, approved, and tasked
-`plan-amend-path-policy-2026-07-20-3c72.md` — 5 tasks/2 phases in
-`tasks-amend-path-policy-c18f.md` (`ready`). Design decision: F001's
-amend-path gap resolved by a **narrow factual-corrections exemption**
-(body-prose fixes to citations/paths/symbols/typos are hand-editable;
-frontmatter/lifecycle state and decisions never are), stated canonically
-in `templates/dot-project-readme.md` and echoed in
-ardd-feedback/ardd-plan prose, `docs/reference/project-files.md`,
-README, CLAUDE.md — explicitly NOT new `--amend` machinery (judgment
-below the deterministic bar). Phase 2: F002 citation-rot sweep — skill
-prose switches to symbol-preferred code citations. No feature slugs
-(`features: []`). Committed to `main` (793d2ed).
-`feedback-badge-workflow-yaml-ff0c.md` deliberately left `open` for a
-separate run. Prior update, same day, `/ardd-implement` coordinator —
-the delegated
-`dot-project-reviewer-guide` worktree run completed all 15 tasks and
-merged (`ort`, no conflicts — merge commit `fbcbbce`, signed); worktree
-reaped; dogfooded install refreshed via `./install.sh .` (the
-home-relative `Source-Path: ~/dev/artifact-driven-dev` and the new
-`.project/README.md` reviewer guide are now live in this repo's own
-install). Shipped: install.sh records `Source-Path` home-relative with
-`~`-expanding readers + legacy-absolute rewrite and history-leak notice
-(leak sweep: nothing beyond Source-Path); `next_step_prompt` enum
-widened to `true|false|auto` (auto-runs a concrete `/ardd-*`
-recommendation; denied prompt = decline) across lint/stamp/skills/docs;
-plans codified as static records (no checkboxes, no derivable counts);
-`templates/dot-project-readme.md` reviewer guide installed to
-`.project/README.md` with an `ardd-version.md` pointer. New
-`test-install-source-path.sh` + CI job. Both features
-(`dot-project-reviewer-guide`, `next-step-prompt-auto`):
-`tasked` → `implemented`. Prior update, same day,
-`/inbox-drain` + `/ardd-feedback` — first live
-run of the new source-side `/inbox-drain` skill (committed `86ee84c`):
-drained one item from `~/dev/.ardd-data/inbox/artifact-driven-dev/`
-into `feedback-badge-workflow-yaml-ff0c.md` (`open`, 1 bug + 3 UX):
-F001 — `templates/ardd-badge-workflow.yml` is invalid YAML (column-0
-heredoc body/terminator ends the `run: |` block scalar at line 53;
-verified fix: indent both to the block scalar's indentation; corrected
-copy exists in moui72/assisted-review, upstreamable); F002 — sweep
-other shipped templates for column-0 heredocs / add a YAML-parse lint;
-F003 — badge workflow needs `branches: [main]` + `workflow_dispatch`
-(currently commits badge JSON onto any triggering branch); F004 —
-consider `[skip ci]` on the sync commit. Inbox file deleted after
-drain; inbox empty. Open feedback now 2 files. Meanwhile the delegated
-`dot-project-reviewer-guide` worktree run advanced to 12/15. Prior
-update, same day, `/ardd-feedback` — drained an out-of-band inbox
-note (`~/dev/.ardd-data/feedback/feedback-1.md`, deleted after drain)
-into `feedback-amend-path-for-skill-written-files-2187.md` (`open`,
-2 UX): F001 — no sanctioned path to amend a skill-written file whose
-content is factually wrong (concrete case: a feedback file's citation
-error with no skill to fix it; generalizes to plans/tasks; resolve by
-either exempting factual corrections from the no-hand-edit rule or
-adding an `--amend`/`ardd-state.sh` path, and state the rule where
-agents read); F002 — skills writing file/line citations write content
-guaranteed to rot; prefer symbol names. The next `/ardd-plan` picks it
-up. Also corrected a stale Feedback-section claim:
-`feedback-repo-critique-6ad1.md` is `planned` (consumed by
-`plan-ardd-state-determinism-2026-07-06.md` on 2026-07-06), not open.
-Meanwhile the delegated `dot-project-reviewer-guide` worktree run is
-in flight at 2/15. Prior update, same day, `/ardd-plan` — consumed both open feedback
-files (`20da`, `19ce` — all 5 items accepted; both now `planned`) and
-targeted `dot-project-reviewer-guide` + `next-step-prompt-auto`:
-drafted, approved, and tasked
-`plan-dot-project-reviewer-guide-2026-07-20-ee87.md` — 15 tasks/4
-phases in `tasks-dot-project-reviewer-guide-b434.md` (`ready`). Phase
-1: portable home-relative `Source-Path` (install.sh writer, `~`-expanding
-readers, legacy rewrite + history advice, leak-class sweep). Phase 2:
-`next_step_prompt` enum widened to `true|false|auto` (lint enum + the
-two prompting skills; denied AskUserQuestion = decline, never
-retry/abort). Phase 3: plans decided as static historical records — no
-checkboxes, no restated derivable counts. Phase 4 (after 3): the
-installed `.project/` reviewer guide + `ardd-version.md` pointer.
-Constitution bumped to v1.12.1 (PATCH — Codex substitution item 4
-covers `auto`). Both features flipped `backlogged→tasked`. Committed to
-`main` (1802b2c). `repo-critique-6ad1` deliberately left open. Prior
-update, same day, `/ardd-feedback`, third pass — vetted the
-"couple next_step_prompt to the harness permission mode" idea and
-logged the recommended shape instead: new backlog entry
-`next-step-prompt-auto` (`backlogged`) — a third enum value `auto`
-that auto-runs the recommended next step when it's a concrete
-`/ardd-*` invocation (explicit and harness-agnostic; the mode-coupled
-variant was rejected: permission mode is only visible via hook
-payloads, would force target-side hook shipping, and
-acceptEdits/auto still show AskUserQuestion) — plus
-`feedback-next-step-prompt-dontask-20da.md` (`open`, 1 UX): dontAsk
-mode denies AskUserQuestion outright, so the two prompting skills
-should state that a denied prompt means decline, never retry/abort.
-Prior update, same day, `/ardd-feedback`, second pass — expanded
-`feedback-source-path-portability-19ce.md` (`open`, now 2 bugs + 1 UX +
-1 Reconsidered) with the full CodeRabbit triage from
-moui72/assisted-review#105: F001 gained design options (tilde form vs
-split file vs drop the field) plus a sweep for other absolute-path
-leaks in committed generated files; new F003 (bug) — /ardd-plan wrote a
-prose count ("four call sites") contradicting its own five-item
-enumeration, nothing reconciles derivable counts; new F004
-(Reconsidered) — plan checkbox staleness: plans look like live mirrors
-but behave as static records, decide which and make it self-evident.
-The fourth finding — a "how to read `.project/`" note for downstream AI
-reviewer configs — was re-filed to the register as
-`dot-project-reviewer-guide` (`backlogged`). Original logging, same
-day: F001 install.sh records `Source-Path: $SCRIPT_DIR` — a
-machine-specific absolute path leaking the username — into the
-committed `.project/ardd-version.md`; F002 `/ardd-update` should detect
-and rewrite the legacy absolute form in consumers and advise on
-git-history repair. The next `/ardd-plan` picks it all up. Prior
-update, same day,
-`/ardd-implement` coordinator — the delegated
-`s1-badge-followups` worktree run completed both tasks and merged
-fast-forward (`e745582`); worktree reaped; badge test re-verified
-27/27 green on `main`. Shipped: install.sh now prints the one-line
-`ARDD_VERSION_BADGE=1` opt-in pointer for no-README targets (F001 —
-new `elif` branch, pointer only, never the snippet), and the
-misdirected-badge advisory's remedy is self-sufficient — replace the
-badge inside the markers with the endpoint form pointing at the
-repo's own `.github/badges/ardd-version.json` — instead of suggesting
-a re-run the reprint guard silences (F002). Reprint guard/coordinate
-fill untouched; red-first (3 new assertions failed before the fix).
-This closes the full S1-sweep loop: brief caught both gaps → feedback
-→ plan → delegated fix → merged. No feature flips (`features: []`).
-Prior update, same day, `/ardd-plan feedback-s1-badge-followups-5e84.md`
-— consumed both S1 follow-up items (file now `planned`): drafted,
-approved, and tasked `plan-s1-badge-followups-2026-07-20-9667.md` — 2
-tasks/2 phases in `tasks-s1-badge-followups-7a62.md` (`ready`).
-Test-first: no-README targets get a one-line ARDD_VERSION_BADGE=1
-opt-in pointer (F001), and the misdirected-badge advisory's remedy
-becomes self-sufficient (replace the badge inside the markers with the
-endpoint form) instead of suggesting a re-run the reprint guard
-silences (F002). Reprint guard/coordinate fill untouched. No feature
-slugs (`features: []`). Committed to `main` (3a14785). Prior update,
-same day, release ops + rename + `/scenario-sweep S1` +
-`/ardd-feedback` — pushed `main` (`4be4848..86f2fec`, 22 commits);
-lint green; beta-release published **v1.0.2-beta.1** (stable remains
-v1.0.1). Renamed the sweep skill: `/prerelease-sweep` →
-`/scenario-sweep` (`a8d3c57`; it outgrew the release-ritual name —
-regression reruns and targeted verification are its main use;
-"-sweep" family with docs-sweep kept; description carries "formerly
-prerelease-sweep" for one release cycle) and `tests/prerelease/` →
-`tests/scenarios/` (inner `scenarios/` flattened; briefs now at
-`tests/scenarios/S<n>.md`). Then ran a targeted S1 sweep (run
-`2026-07-20-a7d7`, ~65K tokens, ~$0.30-0.90, 3m06s) — the first
-exercise of the new badge brief steps against the real `curl | sh`
-`--beta` acquisition of v1.0.2-beta.1: **pass, no release blockers**;
-acquisition/init/lint all clean; all three badge checks (real
-coordinates, no reprint, misdirected-badge advisory) verified live.
-The new brief immediately caught 2 accepted ux findings, logged as
-`feedback-s1-badge-followups-5e84.md` (`open`): F001 the
-ARDD_VERSION_BADGE opt-in mention is gated on README existing, so a
-greenfield new.sh project never sees it; F002 the misdirected-badge
-advisory's suggested remedy is circular (re-run prints nothing under
-the reprint guard). Graduation: n-a for both — the standing brief
-step is what caught them. Prior update, same day, `/ardd-implement`
-coordinator — the delegated
-`sweep-coverage-expansion` worktree run completed all 5 tasks and
-merged fast-forward (`d22d620`); worktree reaped; lint-docs green on
-`main`. Shipped: S1 gained badge steps 6-9 (fake remote, real
-coordinates, no-reprint, opt-in mention, wrong-badge advisory; kept in
-smoke — cheap, no network); S5 gained the bare `/ardd-plan` picker
-step (the unselected feature staying `backlogged` is itself a check);
-S7 gained `--view` with a before/after `shasum` byte-identity check on
-STATUS.md and a full-tier-only `constitution --review` step (S7 has no
-per-step tier convention, so it's a prose gate); S8 gained pre-flight
-Work Queue verdict checks (`independent`, `features: []` → `none`),
-picker matrix annotations, and the in-flight `claimed` check; and the
-prerelease-sweep SKILL.md now carries the graduation mechanism —
-step-3 `never-graduated:` RUN.md lines (grep-and-judge, no script) and
-the step-6 mandatory `graduate ∈ yes/no/n-a` TRIAGE.md column with
-anti-bloat criteria, mirrored in tests/prerelease/README.md. No
-feature flips (`features: []`). The next sweep run exercises the new
-briefs. Prior update, same day, `/ardd-plan feedback-sweep-coverage-expansion-3452.md`
-— consumed the sweep-coverage feedback (all 6 items accepted, file now
-`planned`): drafted, approved, and tasked
-`plan-sweep-coverage-expansion-2026-07-20-d493.md` — 5 tasks/2 phases
-in `tasks-sweep-coverage-expansion-36eb.md` (`ready`). Phase 1
-(parallel-safe, one task per brief): S1 + dynamic-badge steps (against
-the merged 09abad7 behavior), S5 + bare-plan picker, S7 + `--view`
-byte-identical-STATUS check and full-tier `constitution --review`, S8
-+ Work Queue/parallel-matrix verdicts incl. `claimed`. Phase 2: the
-graduation mechanism in prerelease-sweep SKILL.md per the audit
-proposal (design-of-record in the plan) — dispatch-time
-`never-graduated:` RUN.md lines, a mandatory `graduate ∈ yes/no/n-a`
-TRIAGE.md column, feedback-routed brief-coverage items, anti-bloat
-criteria, README mirror; explicitly no new script (fails the
-deterministic bar). No feature slugs targeted (`features: []`).
-Committed to `main` (a217497). Prior update, 2026-07-19,
-`/ardd-implement` coordinator — the delegated
-`dynamic-badge-discoverability` worktree run completed all 4 tasks and
-merged (merge commit `09abad7`, signed with the on-disk key after the
-first attempt hit the locked-1Password signer); worktree reaped;
-dogfooded install refreshed (`8be8b1a`). Shipped: install.sh's badge
-section now fills `OWNER/REPO/BRANCH` in the printed snippet from the
-target's own remote/branch (https + SSH shapes, placeholder + printed
-replace-instruction fallback), guards the `ARDD_VERSION_BADGE=1`
-snippet reprint on existing `ardd-badge-version-start` markers, prints
-a latest-release-badge advisory in both the default and opted-in paths,
-adds the private-repo caveat, and mentions the opt-in in the default
-path's static-badge suggestion; `/ardd-update` now offers the dynamic
-badge at its suggestion-relay step (env var = mechanism, prose =
-interface); templates/USAGE/reference-page docs updated. 8 red-first
-cases added — badge test 23/23 green on `main`, lint-docs clean. No
-feature flips (`features: []`). Separately, the coverage-audit agent
-returned a concrete F006 proposal: graduate dispatcher-stressed
-surfaces into durable briefs via a mandatory `graduate` column in
-sweep triage plus dispatch-time `never-graduated:` advisory lines —
-drafted SKILL.md prose ready for the plan that consumes
-`feedback-sweep-coverage-expansion-3452.md` (now unblocked: its F001
-badge brief depended on this merge). Prior update, same day,
-`/ardd-feedback` — logged
-`feedback-sweep-coverage-expansion-3452.md` (`open`, 6 UX items) from a
-Fable audit of `tests/prerelease/scenarios/` coverage vs recently
-shipped surfaces. Verdict: suite structurally healthy; two expansions
-matter — dynamic-badge checks into S1 (field failure, zero coverage;
-write after the in-flight fix branch merges) and Work Queue/
-parallel-matrix consumer-visible checks into S8 (two real d06a bugs,
-shell-test pin only) — plus smaller S5 (bare-plan picker), S7
-(`--view` single-writer discipline; `constitution --review`)
-extensions and a systemic F006: dispatcher-supplied "recent feature"
-stress ages out after one sweep, so real-finding surfaces should
-graduate into durable briefs. Judged adequate as-is: channels, epics,
-fold/reap/align, rejected/subsumed (deterministic layer), `--slate`,
-source-side skills. Meanwhile the delegated
-`dynamic-badge-discoverability` worktree run is in flight at 3/4.
-Prior update, same day, `/ardd-plan feedback-dynamic-badge-discoverability-a123.md`
-— consumed the badge feedback (all 6 items accepted, file now
-`planned`): drafted, approved, and tasked
-`plan-dynamic-badge-discoverability-2026-07-19-23cf.md` — 4 tasks/3
-phases in `tasks-dynamic-badge-discoverability-6887.md` (`ready`).
-Phase 1 (test-first): install.sh badge-section fixes — fill
-`OWNER/REPO/BRANCH` in the printed snippet from the target's own git
-remote/branch (placeholder + printed replace-instruction fallback),
-marker-guard the `ARDD_VERSION_BADGE=1` snippet reprint, advisory when
-a README carries the wrong latest-release badge shape (both paths),
-private-repo caveat line, and a default-output mention of the
-`ARDD_VERSION_BADGE=1` opt-in. Phase 2: `/ardd-update` offers the
-dynamic badge and sets the env var on the re-run (F006 resolved: env
-var stays the mechanism, skill prose becomes the interface). Phase 3:
-docs/templates (public-repo-only limitation, USAGE routing,
-`ardd-update` reference page). Out of scope deliberately: in-place
-README rewriting (would break the never-edit posture) and the sync
-workflow's push-failure modes. No feature slugs targeted
-(`features: []`). Committed to `main` (721f786). Prior update, same
-day, `/ardd-feedback` — logged
-`feedback-dynamic-badge-discoverability-a123.md` (`open`, 3 bugs + 2 UX
-+ 1 Reconsidered) after the `ARDD_VERSION_BADGE` dynamic-badge feature
-failed in the field: a consuming agent asked to add a version badge
-hand-rolled the wrong shields.io `github/v/release` badge (the exact
-"runs ahead silently" shape the feature was built to prevent) because
-the env-var opt-in is documented nowhere — zero mentions in `skills/`,
-`docs/`, `README.md`, or `USAGE.md`, and `/ardd-update` never offers
-it — then wrapped the wrong badge in `ardd-badge-start/end` markers,
-which now suppress install.sh's suggestion permanently. Also filed:
-the printed two-badge snippet's literal `OWNER/REPO/BRANCH`
-placeholders (the explanatory comment isn't printed), the
-`ARDD_VERSION_BADGE=1` branch reprinting the snippet despite existing
-markers, the missing private-repo caveat (shields.io can't fetch
-`raw.githubusercontent.com` on private repos), and a Reconsidered item
-questioning whether an env-var gate is the right interface for an
-agent-mediated workflow at all. The next `/ardd-plan` run picks it up.
-Prior update, same day, **v1.0.1 shipped** — ship-gate sweep e33f (S3 S5
-S6 S8) came back fully clean: all checklists green, the
-stamp-workflow-mode fix verified live, F002/F003/F004 holding, F001
-reproduced exactly as known-open; only new items were three
-taste-deferred subjective UX notes (advisory-lock same-session noise,
-constitution-tag shared-artifact prevalence, fold-branch leftover).
-Pushed `main` (9fd6fbb), confirmed lint green, dispatched
-`stable-release.yml` (`bump: patch`) through GitHub API 503 flakiness
-— `v1.0.1` published and marked Latest. This closes the loop begun
-with sweep d06a and self-heals its headline finding: `/ardd-update
---stable` consumers now get the fixed install.sh riding the stable
-tag. Full arc today: d06a targeted sweep → 4 fixes → 51a7 regression
-rerun (1 new finding) → stamp-workflow-mode fix → e33f clean gate →
-ship. Prior update, same day, regression sweep 51a7 +
-`stamp-workflow-mode`
-fix — the S3/S5/S6/S8 regression rerun verified all four d06a fixes
-(F002/F003/F004 confirmed live; F001 reproduced as known-open until
-v1.0.1) but surfaced one new bug: `--reconfigure` prose stamps
-`workflow_mode`, which `cmd_stamp`'s enum rejected. Fixed via the full
-loop (feedback `feedback-stamp-workflow-mode-ca7d.md` → plan →
-delegated implement, red-first, 6 new assertions, merged + reaped).
-Per user instruction the ship gate is now: full re-sweep S3 S5 S6 S8;
-on a clean pass, push + dispatch v1.0.1 stable. Prior update, same
-day, `/ardd-implement` coordinator — the delegated
-`sweep-d06a-fixes` worktree run completed all 5 tasks and merged
-fast-forward; worktree reaped; dogfooded install refreshed. Fixed:
-parallel-matrix `features: []` now reads `none` (unknown reserved for a
-broken chain) and the same-file ready-vs-worktree-claim pair gets the
-new `verdict=claimed` (precedence claimed > shared-feature >
-shared-artifact > independent) — red-first, 15/15 asserts green;
-`claimed` named in both consuming skills' prose + reference docs; Work
-Queue data-source clarification applied; and T005 found `c7cb703`
-shipped with no regression test — added case4 (dual-tagged HEAD →
-stable `Source-Ref:`) to `test-install-channel-default.sh`, green
-against the existing fix. Next per user instruction: regression sweep
-S3 S5 S6 S8 re-verifying F001–F004; if clean, push + dispatch v1.0.1
-stable. Prior update, same day,
-`/ardd-plan feedback-prerelease-sweep-d06a-f8ce.md`
-— consumed the sweep feedback (all 4 items incorporated, file now
-`planned`): drafted, approved, and tasked
-`plan-sweep-d06a-fixes-2026-07-19-502e.md` — 5 tasks/3 phases in
-`tasks-sweep-d06a-fixes-ccf0.md` (`ready`). Phase 1 (test-first):
-parallel-matrix verdict fixes — explicit `features: []` reads `none`
-not `unknown`, and a new `verdict=claimed` (precedence claimed >
-shared-feature > shared-artifact > independent) for the same-file
-ready-vs-worktree-claim pair. Phase 2: `claimed` named in both
-consuming skills + the Work Queue data-source clarification
-(tasks-list supplies entries, matrix supplies verdicts) + reference
-docs. Phase 3 [parallel]: verify/add dual-tag stable-preference
-regression coverage for F001, whose operational remedy is dispatching
-v1.0.1 stable. Same-file hard exclusion unchanged. Prior update, same
-day, `/prerelease-sweep S3 S5 S6 S8` + triage +
-`/ardd-feedback` — targeted regression sweep over every post-v1.0.0
-surface (run `2026-07-19-d06a`, ~348K tokens, ~$1.50–3.50, ~8m40s
-parallel). **All four scenarios passed**; the changed surfaces (bare-plan
-picker, Work Queue, parallel-matrix, picker annotations,
-rejected/subsumed, install dual-tag fix) all verified working. 7
-findings triaged → 4 accepted, filed as
-`feedback-prerelease-sweep-d06a-f8ce.md` (`open`, 2 bugs + 2 UX):
-headline F001 — `/ardd-update --stable` on real consumers today
-records a stale beta `Source-Ref:` under `Channel: stable` (v1.0.0's
-install.sh predates fix c7cb703; self-heals when the next stable
-ships, so this argues FOR cutting v1.0.1 promptly); F002
-parallel-matrix `features: []` misread as `unknown`; F003 same-file
-claimed pair mislabeled `shared-feature`; F004 Work Queue prose
-data-source clarification. Deferred/dropped: STATUS.md-accretion
-guidance (taste), fold branch-ref cleanup (taste), advisory-note
-duplicate, harness worktree-binding artifact. Prior update, same day,
-`/ardd-implement` coordinator — the delegated
-`bare-plan-target-prompt` worktree run completed both tasks and merged
-fast-forward; worktree reaped; dogfooded install refreshed. A bare
-`/ardd-plan` now has step 1a: it enumerates plannable inputs
-(backlogged slugs, open feedback, unsurfaced defects) and offers them
-as one multi-select picker forwarding into the existing scope
-machinery; with truly nothing plannable it ends in prose + next-step
-suggestions instead of a complaint. `lint-docs` green;
-`docs/reference/skills/ardd-plan.md` body updated. No bound features
-(plan `features: []`), so no register flip. Prior update, same day,
-`/ardd-plan` (bare) — consumed
-`feedback-bare-plan-target-prompt-dc5f.md` (both UX items accepted, file
-now `planned`): drafted, approved, and tasked
-`plan-bare-plan-target-prompt-2026-07-19-03ba.md` — 2 tasks/2 phases in
-`tasks-bare-plan-target-prompt-6bcd.md` (`ready`). A bare `/ardd-plan`
-gains a step-1a target pick: it enumerates plannable inputs
-(backlogged slugs via `feature-list.sh`, open feedback files,
-unsurfaced defects) and offers them as one multi-select picker,
-forwarding the selection into the existing scope machinery; with truly
-nothing plannable it ends in prose + next-step suggestions
-(`/ardd-backlog`, `/ardd-feedback`, `/ardd-implement`) instead of a
-complaint. No new scripts; `--list`/`--slate`/`--from` untouched. No
-feature slugs targeted. Prior update, same day, `/ardd-implement`
-coordinator — the delegated
-`work-queue-parallel-safety` worktree run completed all 6 tasks and
-merged fast-forward; worktree reaped; dogfooded install refreshed via
-`./install.sh .` (the new script is live in
-`.claude/skills/ardd-scripts/`). Shipped: `scripts/parallel-matrix.sh`
-(pairwise `independent|shared-feature|shared-artifact` verdicts among
-`ready` tasks files and in-flight worktree claims; test-first, 11/11
-regression cases green, CI job + install manifest wired), the Work
-Queue section in `/ardd-status` (report + STATUS.md, omit-if-none), the
-matrix-annotated pick list in `/ardd-implement` (`shared-feature` =
-strong warning, same-file claim stays the only hard exclusion), and
-both reference-doc bodies. Feature `work-queue-parallel-safety`:
-`tasked` → `implemented`. Prior update, same day, `/ardd-feedback` —
-logged
-`feedback-bare-plan-target-prompt-dc5f.md` (2 UX items, `open`): F001 a
-bare `/ardd-plan` should prompt with the plannable items it found
-(backlogged slugs, open feedback, unsurfaced defects) rather than
-complaining about no feedback input; F002 with truly nothing plannable
-it should end in prose + concrete next-step suggestions
-(`/ardd-backlog`/`/ardd-feedback`/`/ardd-implement`). The next
-`/ardd-plan` run picks it up. Prior update, same day,
-`/ardd-plan work-queue-parallel-safety` — drafted,
-approved, and tasked
-`plan-work-queue-parallel-safety-2026-07-19-4c10.md`: 6 tasks/4 phases
-in `tasks-work-queue-parallel-safety-eadb.md` (`ready`). Phase 1
-(test-first): `parallel-matrix.sh` + regression test + CI/install
-wiring — pairwise `independent|shared-feature|shared-artifact` verdicts
-among `ready` tasks files and in-flight worktree claims (feature overlap
-via the tasks→plan→`features:` chain, `unknown` when broken; artifact
-overlap via `[artifacts: ...]` tag intersection; deliberately no path
-heuristics). Phases 2–3: Work Queue section in `/ardd-status` +
-matrix-annotated fan-out picker in `/ardd-implement` (`shared-feature`
-is a strong warning, never a hard exclusion — the same-file claim check
-stays the only hard one; `independent` means "no declared overlap",
-`merge_policy` still governs). Phase 4: docs. Both research-doc open
-questions resolved in the plan; no artifact edits needed. Feature
-`work-queue-parallel-safety`: `backlogged` → `tasked`. Prior update,
-same day, `/ardd-implement` coordinator — the delegated
-`rejected-feature-status` worktree run completed all 8 tasks and merged
-fast-forward; worktree reaped. `rejected` and `subsumed` are now real
-register statuses: `lint-project.sh`'s `FEATURE_STATUS_ENUM` extended;
-`ardd-state.sh feature-flip` grew the five legal transitions
-(`backlogged/planned→rejected`, `backlogged/planned/tasked→subsumed` —
-the asymmetry is deliberate: absorption can be noticed after tasking,
-rejection can't) with regression tests; good-project fixtures, `docs/
-reference/project-files.md`, and `ardd-status`/`ardd-plan` prose all
-updated. Feature `rejected-feature-status`: `tasked` → `implemented`.
-Subagent process note: one commit briefly landed with `--no-verify`
-after the pre-commit suite outran a tool timeout, self-caught and
-redone properly with hooks passing. Prior update, same day,
-`/ardd-research` + `/ardd-backlog` — vetted and
-logged `work-queue-parallel-safety`: a single-pane work-queue view — an
-installed `parallel-matrix.sh` reporting feature- and artifact-overlap
-verdicts among `ready` tasks files and in-flight worktrees (deliberately
-no path heuristics — prose paths are unstructured, so path-contact
-assessment stays agent judgment at presentation time), surfaced as a
-Work Queue section in `/ardd-status` and as annotations on
-`/ardd-implement`'s fan-out multi-select picker. Research doc (lens
-results, rejected alternatives incl. a separate `/ardd-board`
-report-owner skill, open question on whether `shared-feature` should be
-a hard fan-out exclusion or a warning):
-`research-work-queue-parallel-safety-vie-2026-07-19-25f7.md`. Prior
-update, same day, `/ardd-plan rejected-feature-status` — a fable
-design review (dispatched at the approval checkpoint) recommended
-extending the original rejected-only plan to also add `subsumed` (an
-entry whose scope shipped under a *different* feature/plan entry): this
-repo's own register had zero real instances of the `rejected` problem
-but a concrete example of the `subsumed` one, in the feature's own
-`Why:` line. Plan and the already-applied `constitution.md` edit both
-revised in place to cover both statuses (still one MINOR bump, 1.11.1 →
-1.12.0). `subsumed` deliberately not named `superseded` — that word
-already means something different at the *plan* level (a same-document
-replacement for the same feature) than this cross-feature scope-absorption
-case. Approved and tasked: 8 tasks/3 phases in
-`tasks-rejected-feature-status-2fb8.md` (`ready`) — schema
-(`lint-project.sh` enum), state mutation (`ardd-state.sh`'s
-`cmd_feature_flip`, including an asymmetric `tasked -> subsumed` edge
-`rejected` doesn't get, since absorption can be noticed later than a
-pre-work rejection decision), skill prose
-(`ardd-status`/`ardd-plan`/`project-files.md`), and manual verification
-against two throwaway fixtures. Feature `rejected-feature-status`:
-`backlogged` → `tasked`. Prior update, same day, `/ardd-backlog` —
-logged `rejected-feature-status`:
-a new `rejected` status for the feature register's status enum
-(backlogged/planned/tasked/implemented/retired), for a backlogged or
-planned idea the team decides not to pursue and that never gets
-built — a terminal state distinct from `retired` (shipped then
-deliberately removed). Motivated partly by a related, illustrative gap
-from a consumer project: a feature whose scope shipped under a
-*different* plan/feature entry didn't cleanly fit `retired` or
-`implemented` either, and was worked around by closing it as
-`implemented` bound to the other plan with an explanatory note — flagged
-in the entry's `Why:` as a separate, related gap (a "subsumed" status or
-similar) not solved by this same change. Prior update, same day, **v1.0.0
-shipped** — pushed `main` (publishing
-`v0.10.3-beta.4` on the beta channel automatically), confirmed `lint`
-and `beta-release` both green on that commit, then dispatched
-`stable-release.yml` with `bump: major`. The workflow fast-forwarded
-`release` to `main`'s tip and published `v1.0.0` as the latest stable
-release (https://github.com/moui72/artifact-driven-dev/releases/tag/v1.0.0).
-This closes the full loop that began with the 2026-07-18 full prerelease
-sweep: sweep → triage → 6 accepted findings → fix plan → implement →
-regression rerun (all green) → push → CI green → stable dispatch. Prior
-update, same day, `/prerelease-sweep S2 S3 S4 S6 S7` — regression
-rerun (run `2026-07-18-abb8`, ~$2.50–3.75, ~414K tokens, ~8min real
-elapsed) of the scenarios that produced the 6 `prerelease-sweep-fixes`
-findings. **All 6 confirmed fixed** — see
-`dev-notes/prerelease-runs/2026-07-18-abb8/RUN.md`'s regression triage
-table for the full per-finding verification. One incidental catch (not
-a product bug): S6 found this repo's own dogfooded
-`.claude/skills/ardd-scripts/` was several commits stale — fixed
-immediately by re-running `./install.sh .` against this repo itself.
-No new defects found. **Verdict: green light for the 1.0 stable cut**
-as far as this sweep's scope is concerned — the plan/implement/regression
-loop that started with the full sweep is now closed. Prior update, same
-day, `/ardd-implement` — the delegated
-`prerelease-sweep-fixes` worktree run completed all 12 tasks, fixing all
-6 accepted prerelease-sweep findings: `install.sh` now infers `Channel:`
-from `Source-Ref:`'s `-beta.` suffix shape (F001, manually verified
-across all 3 precedence tiers — explicit `ARDD_CHANNEL`, preserved prior
-channel, and the new inferred default); `ARDD_VERSION_BADGE=1` now
-writes its supporting files even when a README already has the static
-badge marker (F002, manually verified — no-README and unset-env-var
-paths confirmed unchanged); `ardd-state.sh feature-flip ...
-implemented` now cross-checks the bound tasks file's completion status
-(F003); `task-check` gives a diagnostic message on a malformed
-colon-suffixed checkbox (F004); `lint-project.sh`'s `plan:`
-path-vs-filename error is now distinct and clear, no more doubled path
-(F005, bad-project fixture's `EXPECTED_BAD_FINDINGS` 38→39); the
-epic-drained-to-zero case is now documented in `ardd-status`'s
-`SKILL.md` (F006). New test coverage:
-`scripts/test-install-channel-default.sh` (new) plus extended cases in
-`test-install-version-badge.sh`, `test-ardd-state.sh`,
-`test-lint-project.sh` — all passing. Merged clean (fast-forward) and
-the worktree reaped. No feature slugs targeted. **Next: regression-rerun
-`/prerelease-sweep S2 S3 S4 S6 S7` (the scenarios that produced these
-findings) before dispatching the stable-release workflow for 1.0.**
-Prior update, same day, `/ardd-plan` — drafted, approved, and tasked
-`plan-prerelease-sweep-fixes-2026-07-18-d341.md`: 12 tasks/3 phases in
-`tasks-prerelease-sweep-fixes-e4d8.md` (`ready`) fixing all 6 findings
-from the full prerelease sweep. Highest priority (F001/T001-T002):
-`install.sh` now infers the default `Channel:` from `Source-Ref:`'s own
-`-beta.` suffix shape instead of hardcoding `stable`, when neither
-`$ARDD_CHANNEL` nor a previously-recorded channel applies — fixes ArDD's
-own installer tripping its own `channel-source-ref-consistency` lint
-check in the normal between-releases state. Also: F002/T003-T005
-(`ARDD_VERSION_BADGE=1` now writes files even if the static badge marker
-already exists), F003/T006-T007 (`feature-flip ... implemented` now
-cross-checks the bound tasks file's actual completion status),
-F004/T008-T009 (`task-check` gives a diagnostic message on a malformed
-colon-suffixed checkbox), F005/T010-T011 (`lint-project.sh`'s `plan:`
-path-vs-filename error no longer garbles the message), F006/T012 (epic-
-drained-to-zero case documented in `ardd-status`'s `SKILL.md`). Consumed
-`feedback-prerelease-sweep-2026-07-18-50ea-ad22.md` (now `planned`). No
-feature slugs targeted. Prior update, same day, `/prerelease-sweep full`
-+ `/ardd-feedback` — ran
-the full 8-scenario (S1–S8) prerelease dry-run in parallel background
-subagents (run `2026-07-18-50ea`, ~$4–7, ~725K tokens, ~12min real
-elapsed) ahead of a potential 1.0 stable cut. 16 raw findings triaged
-(`dev-notes/prerelease-runs/2026-07-18-50ea/TRIAGE.md`) down to 6
-accepted, 2 taste-deferred, 4 duplicates, 4 harness-artifacts (including
-the pre-known, still-unresolved Agent-tool nested-worktree limitation
-from S8 — not consumer-facing). Logged the 6 accepted findings as
-`feedback-prerelease-sweep-2026-07-18-50ea-ad22.md`: **highest priority**
-— `install.sh` defaults `Channel: stable` regardless of the source
-checkout's actual tag shape, so installing from a checkout sitting on a
-beta tag (the normal between-releases state — true right now) produces
-a `Channel:`/`Source-Ref:` pair ArDD's own `channel-source-ref-consistency`
-lint check rejects; reproduced independently in 5 of 8 scenarios. Also:
-`ARDD_VERSION_BADGE=1` silently no-ops when a README already has the
-static badge marker (new-feature gap); `feature-flip ... implemented`
-has no completion cross-check against its bound tasks file; plus 3 minor
-docs/UX findings (`task-check` colon-format error message,
-`plan:` frontmatter path-vs-filename lint error clarity, epic-drained-to-zero
-undocumented in `ardd-status`'s `SKILL.md`). Recommended next: `/ardd-plan`
-these, fix, then regression-rerun `/prerelease-sweep S2 S3 S4 S6 S7`
-before dispatching the stable-release workflow. Prior update, same day,
-`/ardd-implement` — the delegated `docs-drift-fixes`
-worktree run completed all 4 tasks: documented `/ardd-status --view` and
-the register's `epic:` field / by-epic Feature Backlog breakdown on
-`docs/reference/skills/ardd-status.md`, and routed `/ardd-plan --slate`
-in both `USAGE.md`'s "How do I…?" table and `docs/guides/core-loop.md`'s
-narrative — the three gaps `docs-sweep`'s first dogfood run found. No
-feature slugs targeted, so no register flip. Merged clean (fast-forward)
-and the worktree reaped. Prior update, same day, `/ardd-plan` — drafted, approved, and tasked
-`plan-docs-drift-fixes-2026-07-18-2fbb.md`: 4 tasks/2 phases in
-`tasks-docs-drift-fixes-c159.md` (`ready`) fixing the three doc gaps
-`docs-sweep`'s first dogfood run found — `/ardd-status --view` +
-epics/by-epic breakdown missing from `docs/reference/skills/ardd-status.md`,
-and `/ardd-plan --slate` unrouted in `USAGE.md`/`docs/guides/core-loop.md`.
-Consumed both feedback files (both now `planned`). No feature slugs
-targeted — pure doc additions. Prior update, same day, `/ardd-implement`
-— fanned out both `ready` tasks
-files to parallel background worktree subagents (file-disjoint, no
-dependency), both completed and merged clean:
-- `docs-sweep`: all 7 tasks done — new
-  `.claude/skills/docs-sweep/SKILL.md` (local-only, never installed,
-  usage `/docs-sweep [--all]`), plus one-line cross-reference pointers in
-  `CONTRIBUTING.md`'s Releases section and
-  `.claude/skills/prerelease-sweep/SKILL.md`. T007's live dogfood run
-  (scoped to skills changed since `v0.10.2`: `ardd-backlog`, `ardd-plan`,
-  `ardd-refine`, `ardd-status`) found 3 genuine gaps, filed as 2 new open
-  feedback files: `feedback-ardd-status-reference-page-mis-7fa5.md`
-  (`/ardd-status --view` and the `epic:`/by-epic breakdown undocumented
-  on that reference page) and `feedback-ardd-plan-slate-mode-unrouted-c563.md`
-  (`/ardd-plan --slate` documented on its own reference page but unrouted
-  in `USAGE.md`/`docs/guides/core-loop.md`) — matching the research
-  report's pre-cited candidates plus one new find (`--view`).
-  `ardd-backlog --assign-epics` and `ardd-refine constitution --review`
-  were both already fully documented — no findings there. Feature
-  `docs-sweep`: `tasked` → `implemented`.
-- `dynamic-version-badge-sync`: all 6 tasks done — `install.sh` gained
-  an `ARDD_VERSION_BADGE=1` opt-in (mirrors the `ARDD_CHANNEL` env-var
-  pattern) writing `templates/ardd-badge-workflow.yml` (GitHub Action,
-  path-filtered on `.project/ardd-version.md`) and
-  `.github/badges/ardd-version.json` (seeded with the real recorded
-  version) into a target, replacing the static-only badge suggestion
-  with a two-badge pair when opted in; default (unset) path confirmed
-  byte-for-byte unchanged. New `scripts/test-install-version-badge.sh` +
-  CI job added together; caught and fixed a real POSIX-`sh` env-var
-  leak bug in the test itself along the way (a var prefixed onto a
-  shell *function* call persists after return, unlike an external
-  command — fixed with an explicit export/unset wrapper). Feature
-  `dynamic-version-badge-sync`: `tasked` → `implemented`.
-
-Both worktrees merged clean and reaped. Prior update, same day,
-`/ardd-plan docs-sweep dynamic-version-badge-sync`
-— drafted, approved, and tasked as two separate plans (file-disjoint, no
-dependency between them, per an explicit plan-grouping check): `docs-sweep`
-(`plan-docs-sweep-2026-07-18-b6ef.md`, 7 tasks/3 phases in
-`tasks-docs-sweep-e6c1.md`) adds a local-only `.claude/skills/docs-sweep/SKILL.md`
-judging human-facing docs (README/USAGE.md/docs/concepts.md/reference-page
-bodies) against actual SKILL.md behavior, ending in a triage table →
-`/ardd-feedback`, same placement/pattern as `prerelease-sweep`; and
-`dynamic-version-badge-sync` (`plan-dynamic-version-badge-sync-2026-07-18-35aa.md`,
-6 tasks/3 phases in `tasks-dynamic-version-badge-sync-4553.md`) adds an
-`ARDD_VERSION_BADGE=1` env-var opt-in to `install.sh` (mirrors the
-existing `ARDD_CHANNEL` pattern, deliberately not a new CLI flag or
-interactive prompt) writing a shields.io endpoint-badge JSON + sync
-workflow into a target, replacing the current static-only badge
-suggestion with a two-badge pair when opted in. Both features:
-`backlogged` → `tasked`. Prior update, same day, `/ardd-feedback` — re-filed a new-capability item to
-the register as `dynamic-version-badge-sync` (per the skill's mirror
-check: this described something the system doesn't do yet, not a bug/UX
-fix): install.sh installing a dynamic "ArDD version" README badge —
-shields.io endpoint JSON + a GitHub Action watching
-`.project/ardd-version.md` (or firing off `/ardd-update`'s own commit)
-— replacing the current static-only badge suggestion
-(`templates/badge.md`), which structurally can't carry a version since
-it's copied verbatim to every target. Motivated by a consumer
-(assisted-review) already hand-building this exact plumbing from
-scratch. Prior update, same day, `/ardd-implement` — the delegated
-`status-view-mode`
-worktree run completed all 9 tasks: added `/ardd-status --view` (a
-read-only side door running discovery steps 1–5 unchanged, then printing
-the report and stopping — no `STATUS.md` write, no orphaned-flip
-confirmation, no next-step prompt), manually verified against this
-repo's own `.project/` state (T003); wired the three previously-unwired
-migration regression tests (`migration-critique-to-audit`,
-`migration-sync-to-tracker`, `migration-workflow-table`) into
-`.github/workflows/lint.yml`, all three confirmed passing locally first;
-wrote a new `tests/prerelease/scenarios/S8.md` (full tier — Agent-tool
-worktree fan-out delegation, folding in the `fold-to-main.sh`
-eager-background path as setup), extended `S3.md` (channel-switch flow)
-and `S7.md` (`epic:` labels), and registered S8 in
-`.claude/skills/prerelease-sweep/SKILL.md`'s hardcoded `full` tier list
-(`docs/reference/skills/prerelease-sweep.md` doesn't exist, so no doc
-edit needed there). Feature `status-view-mode`: `tasked` → `implemented`.
-Merged clean and the worktree reaped. Also, after a user correction that
-the prior docs-freshness research had drifted off-target (it reframed
-the ask as CI/test-wiring coverage rather than human-facing docs
-staying current), a fresh agent dispatch wrote a corrected report —
-`research-docs-freshness-human-facing-2026-07-18.md` — recommending a
-new local-only `docs-sweep` skill (mirrors `prerelease-sweep`'s
-placement/pattern) judging README/USAGE.md/docs/concepts.md/guides/
-reference-page bodies against actual current SKILL.md behavior, run
-manually before dispatching the stable-release workflow. Cites concrete
-drift already found: epics (`/ardd-status`'s by-epic breakdown,
-`/ardd-tracker`'s milestone mapping) and `/ardd-plan --slate` are
-currently undocumented on the site. Logged as backlog item `docs-sweep`.
-Prior update, same day, `/ardd-plan status-view-mode` — approved and tasked
-`plan-status-view-mode-2026-07-18-ce1f.md`: 9 tasks/3 phases in
-`tasks-status-view-mode-6377.md` (`ready`) covering a read-only
-`/ardd-status --view` mode (Phase 1), the CI-wiring fix for the three
-unwired migration tests (Phase 2), and prerelease-sweep scenario
-additions — new S8 (Agent-tool fan-out delegation), extended S3 (channel
-switching), extended S7 (`epic:` labels) (Phase 3). Consumed both
-feedback batches from earlier today
-(`feedback-ci-migration-tests-unwired-37ee.md`,
-`feedback-prerelease-sweep-scenario-gaps-95f6.md`, both now `planned`).
-Feature `status-view-mode`: `backlogged` → `tasked`. Prior update, same
-day, `/ardd-implement` — the delegated `constitution-trim-review-relev`
-worktree run completed all 6 tasks: added
-a `--review` mode to `/ardd-refine constitution` (enumerate declared
-principles → ground a keep/trim judgment per principle against the
-current project → batch-confirm trim candidates, mirroring `/ardd-plan`
-step 3c's shape → apply confirmed trims via the skill's existing
-constitution special-case handling — version bump, Sync Impact Report →
-report), documented in `docs/reference/skills/ardd-refine.md` and
-`USAGE.md`, and dogfood-verified live against this repo's own real
-`constitution.md` (T006): all nine current principles graded **keep**,
-zero trim-candidates — expected, since this constitution was recently
-pruned. No write was made to the real constitution.md by the dry run.
-Feature `constitution-trim-review-relev`: `tasked` → `implemented`.
-Merged clean (`ort` strategy, no conflicts) and the worktree reaped. Also,
-`/ardd-plan status-view-mode` drafted `plan-status-view-mode-2026-07-18-ce1f.md`
-consuming both feedback batches from earlier today (`feedback-ci-migration-tests-unwired-37ee.md`,
-`feedback-prerelease-sweep-scenario-gaps-95f6.md`, both now `planned`) plus
-the `status-view-mode` backlog item — awaiting approval at the
-checkpoint. Separately, an `/ardd-research`-shaped agent dispatch wrote
-`research-docs-freshness-skill-2026-07-18.md`: recommends a new
-deterministic `scripts/lint-coverage.sh` (CI-wired, catches
-tests-not-wired-into-CI drift, complementary sibling to `lint-docs.sh`)
-plus a new source-side `coverage-sweep` skill (judgment half, modeled on
-`prerelease-sweep`) — not yet actioned into a plan. Prior update, same
-day, `/ardd-backlog` — logged `status-view-mode`: a
-`/ardd-status --view` read-only mode that reports a summary,
-incomplete/in-flight snapshot, and recommended next step without
-regenerating `STATUS.md` (mirrors `/ardd-plan --list` /
-`/ardd-implement --list`'s read-only side-door pattern). Also, a review
-agent weighed the `constitution-trim-review-relev` design question
-(audit vs. refine) and confirmed the plan's existing choice —
-`/ardd-refine --review` — is correct: single-writer ownership of
-constitution.md edits and the skill-naming system both point there, not
-at `/ardd-audit` (whose job is surface-only, into `audit.md`); no plan
-change needed. Prior update, same day, `/ardd-implement` — the delegated
-`backlog-assign-epics-automated` worktree run reported back with all 6
-tasks complete: T001 (test-first) added a red case to
-`scripts/test-ardd-state.sh` for `feature-field <slug> epic <value>`;
-T002 added `epic` to `ardd-state.sh`'s `feature-field` valid-key case
-statement; T003–T005 added the `--assign-epics` mode (enumerate →
-propose thematic groupings → batched confirm → apply via `feature-field
-... epic` → report) to `skills/ardd-backlog/SKILL.md`; T006 documented it
-in `docs/reference/skills/ardd-backlog.md`. Feature
-`backlog-assign-epics-automated`: `tasked` → `implemented`. Merged
-clean (`ort` strategy, no conflicts) and the worktree reaped. Prior
-update, same day, `/ardd-plan constitution-trim-review-relev` — drafted,
-approved, and tasked `plan-constitution-trim-review-relev-2026-07-18-8c82.md`
-(solo mode, no branch gate, on `main`): adds a `--review` mode to
-`/ardd-refine constitution` that enumerates the declared principles,
-grounds a keep/trim-candidate judgment for each against the current
-project (never a fixed rubric), batch-confirms proposed trims (mirroring
-`/ardd-plan` step 3c's shape), then reuses the skill's existing
-constitution-special-case handling (version bump, Sync Impact Report) to
-apply confirmed removals. 6 tasks across 3 phases in
-`tasks-constitution-trim-review-relev-3a39.md` (`ready`). No artifact
-changes (this is a new skill capability, not a constitution change).
-Feature `constitution-trim-review-relev`: `backlogged` → `tasked`.
-Meanwhile, `/ardd-implement` delegated
-`tasks-backlog-assign-epics-automated-e23f.md` to a background worktree
-subagent (`delegation: eager`), now at 5/6 tasks — not yet merged. Prior
-update, same day, `/ardd-implement` + `/ardd-plan` — delegated
-worktree run completed and merged both tasks of
-`tasks-channel-source-ref-consistency-824e.md` (now `completed`): T001
-(test-first) added `Channel:`/`Source-Ref:` fixtures to
-`good-project`/`bad-project` (consistent vs. atelier-shaped mismatch,
-`EXPECTED_BAD_FINDINGS` 37→38) plus a targeted message-quality case;
-T002 added the actual check to `lint-project.sh` — reuses
-`install.sh`/`ardd-update-check.sh`'s existing `sed` read style for
-`.project/ardd-version.md` and `source-resolve.sh`/`next-version.sh`'s
-`-beta.` prerelease-suffix recognition, reporting `Channel: stable but
-Source-Ref: <value> is a prerelease tag — a prerelease tag under a
-stable channel is self-contradictory` when triggered. Feature
-`channel-source-ref-consistency`: `tasked` → `implemented`. Merged
-fast-forward (`9ce5b1a..3447331`) and the worktree reaped. Meanwhile,
-also drafted, approved, and tasked
-`plan-backlog-assign-epics-automated-2026-07-18-3d8f.md` (the second
-parallel-set item from today's `--slate` run, solo mode, no branch
-gate, on `main`): adds a new `epic` key to `ardd-state.sh
-feature-field` (Phase 1, test-first) and a new `/ardd-backlog
---assign-epics` sweep mode (Phase 2) structurally mirroring the
-existing `--from-artifacts` bulk mode — walk the register for
-`epic`-less entries, propose thematic groupings by judgment (never
-forcing a non-clustering item into a bucket), one batched
-`AskUserQuestion` confirmation, apply accepted groups via the new write
-path. 6 tasks across 2 phases in
-`tasks-backlog-assign-epics-automated-e23f.md` (`ready`). No artifact
-changes (no new principle/workflow field — this is a new skill
-capability, not a constitution change). Prior update, same day,
-`/ardd-plan channel-source-ref-consistency` —
-drafted, approved, and tasked
-`plan-channel-source-ref-consistency-2026-07-18-461b.md` (solo mode, no
-branch gate, on `main`). Picked as the top-priority item from an
-`/ardd-plan --slate` run over the 4-item backlog: graded
-`channel-source-ref-consistency` `high` confidence (a clear, well-
-understood seam — `install.sh:371-388` writes `Channel:`/`Source-Ref:`,
-`lint-project.sh` has zero existing check for their consistency,
-confirmed via grep), pairwise file-disjoint from the other three
-backlogged items (no bundles this round). Also graded
-`backlog-assign-epics-automated` and `constitution-trim-review-relev`
-`medium` (parallel-set-safe) and `codex-second-harness-support`
-solo-deferred (gated on its own drafted plan's blocking live-CLI
-smoke-test go/no-go, not a parallel-set candidate at this time). No
-artifact changes needed (no new principle/workflow field). 2 tasks in
-1 phase in `tasks-channel-source-ref-consistency-824e.md` (`ready`):
-T001 (test-first) adds consistent/inconsistent
-`.project/ardd-version.md` fixtures to `good-project`/`bad-project`
-plus a targeted message-quality case; T002 adds the actual
-`Channel:`/`Source-Ref:` consistency check to `lint-project.sh`,
-reusing `source-resolve.sh`/`next-version.sh`'s existing prerelease-tag
-recognition. Feature `channel-source-ref-consistency`: `backlogged` →
-`planned` → `tasked`. Prior update, same day, `/ardd-update` — this
-repo's own installed
-`.claude/skills/ardd-scripts/` copy had gone stale (last self-installed
-at commit `a656512`, predating the `feature-list.sh` manifest-gap fix
-and several other recent scripts), discovered when `/ardd-backlog
---list` (not a real mode — corrected to a direct `feature-list.sh` call)
-hit a missing-file error. Re-ran `install.sh` against this repo
-(self-hosted, `channel=dev`); refreshed to commit `66bd323`
-(today's docs-audit commit). No migrations pending, no workflow-field
-backfill needed (all four already set), `.worktreeinclude`/
-`.gitattributes` already correct. Prior update, same day, release ops —
-pushed `main` (`a802dbc..db95bae`,
-publishing the v0.10.2 fix batch as `v0.10.2-beta.1`), then ran a
-targeted `/prerelease-sweep S1 S2 S3` regression check (run
-`2026-07-17-db69`) specifically re-verifying S2-F001 (`ardd-init` diff
-guidance) and S3-F001 (`plan_preview` gating) via live reproduction —
-both confirmed genuinely fixed, no new findings, no branch-protection
-snag this time. Dispatched `stable-release.yml` (`bump: patch`) —
-succeeded on the first try. **`v0.10.2` is now the published stable
-release**; confirmed via direct `curl` that the diff-verification
-guidance is live in the real `release` branch. This closes the
-phantom-completion incident end-to-end: found by the full sweep →
-fixed inline with explicit diff verification per task → regression-
-checked → shipped. Prior update, same day, `/ardd-implement` — ran
-`tasks-v0-10-2-fixes-53cb.md`
-**inline, deliberately not delegated** (given the tasks file exists
-because a delegated subagent falsely claimed task completion, this run
-verified each task's `git diff` in real time rather than trusting
-another delegated self-report). All 7 tasks completed (now
-`completed`): T001-T003 re-landed the three missing v0.10.1 edits —
-`ardd-status` now handles `dev-ahead` distinctly from `behind`,
-`ardd-init` step 7 warns that a `feat:` commit's message isn't proof of
-its diff, `ardd-plan` step 10 actually gates the browser-preview
-question on `plan_preview` — each verified via `git diff` against the
-named file before being marked complete. T004 (test-first, `--no-verify`
-documented in the commit body per this repo's shell-test convention,
-since there's no language-level xfail marker) added a regression case
-proving `tasks-flip completed` didn't check for unchecked tasks; T005
-fixed `ardd-state.sh` to refuse the flip with the still-open task ID(s)
-named, confirmed green. T006 added a mandatory pre-`task-check` diff
-self-check to `/ardd-implement` step 7 itself — the process fix meant to
-prevent this exact failure class going forward. T007 confirmed the full
-test suite green and the tree clean before completion. No feature flip
-— `features: []`. All work committed directly to `main` (solo mode, no
-worktree). Prior update, same day, `/ardd-plan` — drafted, approved, and
-tasked
-`plan-v0-10-2-fixes-2026-07-17-4465.md` (solo mode, no branch gate, on
-`main`). Consumed `feedback-prerelease-full-sweep-v0-10-1-e5d8.md`'s all
-3 items (accepted): F001 (the phantom-completion gap — the three missing
-edits actually need to land this time, each with an explicit
-`git diff`-based self-check before being marked complete), F002
-(`ardd-state.sh tasks-flip completed` should refuse when unchecked tasks
-remain), F003 (the process root cause — `/ardd-implement` step 8 needs a
-mandatory pre-`task-check` diff self-check). No unsurfaced defects, no
-feature slugs targeted. 7 tasks across 4 phases in
-`tasks-v0-10-2-fixes-53cb.md` (`ready`): Phase 1 (T001-T003, parallel)
-re-lands the three missing v0.10.1 edits, each carrying an explicit
-"verify via `git diff` before marking complete" instruction naming that
-this exact task was claimed-but-unapplied once already. Phase 2 (T004,
-test-first; T005) adds a real deterministic `tasks-flip` checkbox
-refusal. Phase 3 (T006) tightens `/ardd-implement`'s own process so this
-failure mode is caught before a commit ever claims work that wasn't
-done. Phase 4 (T007) is the pre-cut verification gate; the plan's own
-Phase 4 notes (push, targeted `/prerelease-sweep S1 S2 S3` regression,
-`stable-release` dispatch) are release ops for after tasks complete, not
-tasked directly. Prior update, same day, `/ardd-feedback` — a
-`/prerelease-sweep full` run
-against **v0.10.1** (run `2026-07-17-b924`, S1-S7, all completed
-cleanly) found a critical issue: three tasks from the earlier
-`tasks-feedback-batch-ec6e.md` run (T006, T007, T010) were marked `[x]`
-complete with commits describing the intended edit, but the edits never
-actually landed — independently reproduced by 4 of 7 scenarios (S2, S3,
-S5, S7). **v0.10.1, already stable, ships this gap**: `ardd-status`
-still doesn't handle the `dev-ahead` update-check outcome, `ardd-init`
-step 7 still has no diff-verification guidance, and `plan_preview` is
-fully wired everywhere except the one place (`ardd-plan` step 10) that's
-supposed to consume it. Logged as
-`feedback-prerelease-full-sweep-v0-10-1-e5d8.md` F001, alongside F002
-(`ardd-state.sh tasks-flip completed` doesn't verify checkboxes, and
-`completed`'s terminal design leaves no scripted recovery from a
-wrongly-flipped file — found by S6) and F003 (the process root cause:
-`/ardd-implement`'s task-completion flow has no verification step
-before `task-check` marks a task done). Not yet consumed by a plan.
-Everything else in the sweep was clean or non-actionable (S1/S4 clean
-passes; S1-F001/S4-F001 working-as-designed/harness-only). Cost: ~649K
-tokens / ~$3-10 across 7 subagents; full breakdown in
-`dev-notes/prerelease-runs/2026-07-17-b924/RUN.md`. Prior update, same
-day, release ops — pushed `main` (`a978ab7..a802dbc`,
-publishing the `new.sh` fix as a beta), then ran a `/prerelease-sweep S1`
-regression check (run `2026-07-17-cc3b`) against the real quickstart
-URL. It confirmed the fix is correct on `main` but caught something more
-important: `new.sh`'s public curl URL always resolves through the
-`release` branch regardless of `--beta`/`--stable`, and `release` was
-several commits behind — so the fix hadn't reached real users yet.
-Dispatching `stable-release.yml` then failed
-(`GH006: Protected branch update failed — must not contain merge
-commits`): `main`'s history holds two legitimate non-fast-forward merge
-commits from 2026-07-15 (sanctioned by `merge_policy: auto`), colliding
-with `release`'s `required_linear_history` GitHub branch-protection rule
-— a latent conflict between two of this repo's own standing decisions,
-not a new bug, but one that would have silently blocked every future
-stable dispatch. With explicit user permission, disabled just that one
-protection flag via `gh api` (nothing else touched — force-pushes and
-deletions on `release` are still blocked) and re-dispatched successfully.
-**`v0.10.1` is now the published stable release**; `release` is
-fast-forwarded to `main`'s tip; confirmed via a direct `curl` that the
-fixed `new.sh` guard logic is live at the real public URL. Prior update,
-same day, `/ardd-implement` — delegated worktree run
-completed and merged all 12 tasks of `tasks-feedback-batch-ec6e.md`
-(now `completed`): T001/T002 (test-first) fixed the release-blocking
-`new.sh` git-init isolation bug — `new.sh:242` now compares
-`git rev-parse --show-toplevel` against `$TARGET`'s realpath instead of
-using `--is-inside-work-tree` (true for any nested directory, not just
-a real repo root), so a target nested under an existing git-controlled
-directory finally gets its own isolated repo instead of silently
-inheriting the outer one's identity. T003/T004 fixed the two downstream
-`install.sh` symptoms: gitignore-diagnostic misattribution (now confirms
-`$TARGET` is its own repo top-level before trusting `check-ignore`) and
-`.project/.lock` now gets written to `.gitignore` unconditionally
-instead of only mentioned in printed text. T005 taught
-`ardd-update-check.sh` to distinguish a dev-mode checkout genuinely
-*behind* a release tag from one that's *ahead* with no `Source-Ref`
-(new outcome token: `dev-ahead`); T006 made `/ardd-status`'s banner
-treat that distinctly (never recommends `/ardd-update` when it would
-regress the target). T007 added diff-verification guidance to
-`/ardd-init` step 7's git-log feature-extraction heuristic. T008/T009
-(test-first) added the new `plan_preview` workflow field
-(`always-browser`/`always-console`/`ask`) to `lint-project.sh`'s enum
-and `ardd-state.sh stamp`, mirroring `delegation`/`merge_policy`; T010
-wired `/ardd-plan` step 10's browser-preview question to respect it;
-T011 generalized the constitution's Governance "Exception" clause to
-cover any workflow-field-enum member by reference (closing the
-pre-existing gap where `delegation`/`merge_policy`/
-`update_check_max_age_days` were already exempt in practice but
-unnamed) — constitution PATCH-bumped 1.11.0 → 1.11.1, new SIR entry
-prepended correctly (avoided the bare-`---`-inside-a-comment parsing
-trap this same plan's drafting session hit and reverted earlier
-today). T012 documented `plan_preview` in the `/ardd-plan` reference
-doc. Full test suite confirmed passing before merge. No feature flip —
-this plan consumed feedback only (`features: []`). Merged fast-forward
-(`3980965..4fc55d8`) and the worktree reaped. Prior update, same day,
-`/ardd-plan` — drafted, approved, and tasked
-`plan-feedback-batch-2026-07-17-e977.md` (solo mode, no branch gate, on
-`main`). Consumed all 3 open feedback files (5 items — all accepted):
-`feedback-plan-preview-setting-63b3.md` F001 (Reconsidered — the
-`/ardd-plan` browser-preview prompt becomes a configurable
-`plan_preview` workflow field), `feedback-prerelease-full-sweep-62ae.md`
-F001/F002 (`.project/.lock` gitignore gap; `ardd-init`'s unverified
-git-log trust gap), `feedback-prerelease-smoke-sweep-849d.md` F001/F002
-(`new.sh` git-init isolation bug + its `install.sh` downstream symptom;
-misleading dev-mode "behind" wording) — all now `planned`. During the
-Reconsidered-item negotiation, also surfaced and confirmed a related fix:
-the constitution's Governance "Exception" clause named only
-`workflow_mode`/`next_step_prompt` as SIR-exempt workflow fields, though
-`delegation`/`merge_policy`/`update_check_max_age_days` were already
-exempt in practice — generalizing that clause to cover the enum by
-reference is now T011, rather than applied ad hoc during planning (an
-earlier direct edit was made and then correctly reverted, since a
-Reconsidered item's artifact fix belongs in the plan as a task, not
-applied during `/ardd-plan` itself). No unsurfaced defects, no feature
-slugs targeted (`features: []`). 12 tasks across 3 phases in
-`tasks-feedback-batch-ec6e.md` (`ready`): Phase 1 (T001, test-first;
-T002; T003, T004 parallel) fixes `new.sh`'s git-init guard and its two
-downstream `install.sh` symptoms (gitignore misattribution, `.lock` not
-gitignored). Phase 2 (T005, T006; T007 parallel) fixes
-`ardd-update-check.sh`'s misleading "behind" wording for ahead-of-tag
-dev-mode installs and adds diff-verification guidance to `ardd-init`
-step 7. Phase 3 (T008, test-first; T009; T010 parallel; T011
-`[artifacts: constitution]`; T012) adds the `plan_preview` workflow
-field end-to-end (enum, `ardd-state.sh stamp`, `ardd-plan` step 10
-consumption, the Governance clause fix, and docs). Open questions: exact
-naming for `ardd-update-check.sh`'s new ahead-of-tag outcome token; and
-whether `plan_preview: always-console` should stay silent across a
-Revise loop within the same run (leaning yes). Prior update, same day,
-`/ardd-feedback` — logged
-`feedback-plan-preview-setting-63b3.md`, 1 Reconsidered item: the
-`/ardd-plan` approval checkpoint's "view in browser?" prompt (added
-`62052ae`) always asks on every run; reconsidered to a configurable
-workflow field (`always-browser` / `always-console` / `ask`, default
-`ask` = current behavior), analogous to `delegation`/`merge_policy`/
-`next_step_prompt`. Not yet consumed by a plan. Prior update, same day,
-`/ardd-backlog` — logged
-`constitution-trim-review-relev`: an agent-driven review mode (e.g.
-`/ardd-refine constitution --review`) that audits a project's existing
-constitution principle-by-principle and proposes trimming any that
-aren't relevant or load-bearing for guiding agents toward better code
-quality on that project, batched for confirmation. Why: constitutions
-only ever grow via `/ardd-init`'s suggestion catalog and later
-`/ardd-refine` passes — nothing currently looks back and asks whether an
-accumulated principle still earns its place. Prior update, same day,
-`/ardd-feedback` — a `/prerelease-sweep full` run
-(2026-07-17-fab5, S1-S7) completed, and its S6 result triggered a
-redesign: S6 (delegated worktree execution) was replaced with a
-script-layer-only scenario (`tests/prerelease/scenarios/S6.md`) that
-verifies `worktree-align.sh`/merge/`worktree-reap.sh` against a
-manually-created second worktree, dropping the untestable-from-sweeps
-`Agent`-tool nested-worktree question — that question now has an
-actionable manual recipe in the README's Coverage backlog instead
-(second terminal, scratch project, real `/ardd-implement` delegate).
-Feedback captured: `feedback-prerelease-full-sweep-62ae.md` (2 new
-items — F001 bug, `install.sh` never actually gitignores
-`.project/.lock`, only mentions it in printed text; F002 ux,
-`/ardd-init` step 7 trusts `feat:` commit titles without diff
-verification, risking a false "implemented" register entry on repos
-with misleadingly-labeled commits). `feedback-prerelease-smoke-sweep-849d.md`
-(still open) got two reconfirmation notes appended, not new items: its
-F001 (`new.sh` git-init nesting bug) reproduced again plus a new
-downstream symptom (`install.sh`'s gitignore-diagnostic misattribution);
-its F002 (misleading "behind" wording) reproduced again on a different
-dev-mode install. Full sweep found zero net-new bugs beyond those two —
-S3 (`--slate` against atelier's real 12-item backlog) and S5 (solo
-inline core loop) both passed clean. Cost: ~615K tokens / ~$3-9 across
-the 6 background subagents (S6 ran inline in the dispatching session);
-full breakdown in `dev-notes/prerelease-runs/2026-07-17-fab5/RUN.md`
-(gitignored). Prior update, same day, `/ardd-backlog` — logged
-`backlog-assign-epics-automated`: an automated `/ardd-backlog
---assign-epics` pass over the feature register, open feedback, and
-`DEFECTS.md` that proposes `epic:` groupings for related items, batched
-for confirmation. Why: `epics-grouping-in-feature-regi` added the field
-and manual by-epic views, but assignment itself is still entirely
-manual; conceptually the epic-dimension counterpart to `--slate`'s
-footprint-based grouping. A `/prerelease-sweep full` run is in progress
-concurrently (run `2026-07-17-fab5` — S1-S3, S5, S6 completed; S4 still
-running); this entry was logged mid-sweep and doesn't touch any sweep
-state. Prior update, same day, `/ardd-feedback` — captured the prerelease smoke
-sweep's (run `2026-07-17-1d42`) 2 accepted feedback items in
-`feedback-prerelease-smoke-sweep-849d.md`: F001 (bug) `new.sh`'s git-init
-guard walks up the directory tree for any enclosing `.git`, silently
-folding a nested target into the outer repo instead of creating its own
-— a release-blocker candidate, not scratch-harness-specific; F002 (ux)
-`ardd-update-check.sh`/`/ardd-status`'s "behind" wording is misleading
-for a dev-mode checkout that's ahead of the latest release tag. The
-third accepted finding (S7-F002 — nothing validates `ardd-version.md`'s
-`Channel:`/`Source-Ref:` consistency) was re-filed as a new backlog
-feature instead, `channel-source-ref-consistency` — it's a missing
-capability, not a bug in existing behavior. Prior update, same day,
-`/ardd-implement` — delegated worktree run completed
-and merged all 9 tasks of `tasks-plan-time-defrag-slate-analysi-2c40.md`
-(now `completed`): added `/ardd-plan --slate` — a read-only advisory mode
-that computes an ephemeral "defrag" grouping over the open backlog
-(bundles / parallel sets / solo-deferred, per the two research
-prototypes' two-axis overlap+dependency model) and ends in a recommended
-`/ardd-plan <slug...>` invocation, never writing a plan/tasks file or
-touching the register itself. T008's end-to-end dry-run against this
-repo's own real backlog (N=1 at the time — only
-`codex-second-harness-support` was `backlogged`) correctly exercised the
-N=1 branch. Merged fast-forward (`320f7c8..a6ca315`) and the worktree
-reaped. Feature `plan-time-defrag-slate-analysi`: `tasked` →
-`implemented`. Prior update, same day, a `/prerelease-sweep smoke` run
-(S1, S5, S7; run `2026-07-17-1d42`, reports in `dev-notes/` — gitignored,
-not committed) found one likely release-blocking bug: `new.sh`'s
-git-init guard (`new.sh:240`, `git rev-parse --is-inside-work-tree`)
-walks *up* the directory tree for any enclosing `.git`, so a target
-nested under an existing git-controlled directory silently skips
-`git init` and gets folded into the outer repo (shares its `.git`,
-branch, history, remote) — not scratch-harness-specific, hits any real
-user whose target sits under a monorepo/dotfiles/versioned home dir.
-Also found: `ardd-update-check.sh` reports "behind" for a dev-mode
-checkout that's actually ahead of the latest tag (misleading — the
-suggested fix would regress it), and a pre-existing atelier-repo
-`Channel`/`Source-Ref` inconsistency nothing currently detects. Triage
-table at `dev-notes/prerelease-runs/2026-07-17-1d42/TRIAGE.md`; not yet
-folded into `/ardd-feedback` — awaiting user go-ahead. Prior update,
-same day, `/ardd-plan plan-time-defrag-slate-analysi` —
-drafted, approved, and tasked
-`plan-plan-time-defrag-slate-analysi-2026-07-17-1a95.md` (solo mode, no
-branch gate, on `main`). Designs a codified `/ardd-plan --slate` mode,
-grounded in both prototype `/ardd-research` passes' findings
-(sync-tab-scroll N=1: the N≤1 degenerate branch; atelier N=12: the
-two-axis file-overlap/dependency model, confidence grading, and the
-register-direct-read discipline). No new deterministic script — footprint
-grading is agent judgment (Principle II only mechanizes what's actually a
-pure function of file state); the one mechanical sub-step
-(enumerate `backlogged` items) reuses the existing
-`scripts/feature-list.sh --status backlogged`, unmodified. No artifact
-changes (this repo has only `constitution.md`, and the feature introduces
-no new principle). 9 tasks across 4 phases in
-`tasks-plan-time-defrag-slate-analysi-2c40.md` (`ready`): Phase 1
-(T001–T002) adds the `--slate` argument-shape dispatch and the N=0/N=1
-degenerate branch; Phase 2 (T003–T005) adds per-item confidence grading
-and the two-axis pairwise relation-finding, verified against a scratch
-fixture; Phase 3 (T006–T008) adds classification into bundle/parallel-
-set/solo-deferred plus the `next_step_prompt` handoff, verified
-end-to-end against this repo's own real 2-item backlog; Phase 4 (T009)
-updates the `/ardd-plan` reference doc. No test tasks for the skill-prose
-edits themselves (matches the `--list` precedent — read-only, no state
-mutation to regression-test). `plan-time-defrag-slate-analysi`:
-`backlogged` → `planned` → `tasked`. The plan's own draft initially
-omitted the required Phase Breakdown section — caught and fixed before
-tasking, re-confirmed with the user. Prior update, same day (2026-07-16),
-`/ardd-implement` — delegated worktree run completed
-and merged all 4 tasks of `tasks-delegation-preflight-autocommit-a977.md`
-(now `completed`): T001 (audit) built a throwaway scratch repo and tested
-`skills/ardd-implement/SKILL.md` step 3's pre-flight check against 3
-scenarios — untracked plan file and modified tasks file both correctly
-surfaced, but a `plan:` frontmatter naming a nonexistent plan file was a
-real scope-miss (`git status --short` on a missing path prints nothing
-and exits 0, indistinguishable from clean). T002-T003 changed the
-pre-flight to auto-commit the plan/tasks file(s) in solo mode (scoped
-`git add`, signed commit per this repo's `CLAUDE.md` convention,
-announcing paths + short hash) and moved it ahead of the `on_default:
-false` → `fold-to-main.sh` step so an uncommitted file no longer causes
-a spurious fold refusal. T004 fixed the scope-miss found in T001: the
-check now verifies the resolved plan file exists on disk before running
-`git status`, stopping and surfacing a message (both modes) if it's
-missing. No feature flip — this plan consumed feedback only
-(`features: []`), a documented no-op case. Merged fast-forward
-(`f3f81db..55571a3`) and the worktree reaped. Prior update, same day,
-`/ardd-plan` — drafted, approved, and tasked
-`plan-delegation-preflight-autocommit-2026-07-16-0ca8.md` (solo mode, no
-branch gate, on `main`). Consumed the sole open feedback item
-(`feedback-delegation-preflight-autocommit-06b1.md` F001 — user chose to
-accept it, over declining, after the plan surfaced that the root cause
-was unconfirmed; file now `planned`). No unsurfaced defects, no feature
-slugs targeted, no artifact changes (pure skill-prose fix). 4 tasks
-across 3 phases in `tasks-delegation-preflight-autocommit-a977.md`
-(`ready`): Phase 1 (T001) manually audits the existing
-`skills/ardd-implement/SKILL.md` step 3 pre-flight check's frontmatter
-resolution and `git status --short` scoping against three scenarios,
-since the feedback's root cause couldn't be confirmed and the check may
-have a latent scope-miss independent of the UX question. Phase 2 (T002;
-T003, parallel) changes the check to auto-commit the plan/tasks file(s)
-in solo mode instead of asking, and reorders it ahead of the
-`on_default: false` → `fold-to-main.sh` dirty-tree check so an
-uncommitted file doesn't cause a spurious fold refusal. Phase 3 (T004)
-is conditional — fixes any mechanics gap Phase 1 finds, or is skipped
-with a recorded no-defect outcome. Open question: whether the
-auto-commit message should embed the tasks file's slug for
-traceability (leaning yes). Prior update, same day, `/ardd-feedback` —
-logged
-`feedback-delegation-preflight-autocommit-06b1.md`, 1 Reconsidered item:
-the delegation pre-flight check added for
-`feedback-uncommitted-plan-tasks-delegat-a3ff.md` (`skills/ardd-implement/SKILL.md`
-step 3) currently offers to commit an uncommitted plan/tasks file or
-blocks delegation — user reports this recurs often enough ("this
-happens often") that the ask-first behavior should be reconsidered to
-an automatic commit in solo mode, since it's almost always the obvious
-right move (the files are what the immediately-prior `/ardd-plan` run
-just wrote). Not yet consumed by a plan. Prior update, same day,
-`/ardd-implement` — delegated worktree run completed
-and merged all 5 tasks of `tasks-install-manifest-gap-fix-6cfb.md` (now
-`completed`): T001/T002 (test-first, one commit — T001's assertion is
-expected-red and the pre-commit hook runs the full `test-*.sh` suite)
-added a `feature-list.sh` executable check to
-`test-install-worktreeinclude.sh`'s Case 1b block, confirmed it failed
-against the real bug, then added the missing `cp`/`chmod +x`/`echo`
-lines to `install.sh` (the script was absent from all three, not just
-the `cp` list as first suspected). T003/T004 added
-`scripts/test-install-manifest-complete.sh` — diffs scripts referenced
-by `skills/*/SKILL.md` and `install.sh`'s own `chmod +x` list against
-its actual `cp` manifest; the subagent caught and fixed a real bug in
-its own new script during development (a newline- vs space-separated
-matching mismatch that made every script falsely report missing) before
-committing. T005 added the matching CI job. Merged fast-forward
-(`83a0da2..209b339`, no signing needed) and the worktree reaped. No
-feature flip — this plan consumed feedback only (`features: []`), a
-documented no-op case. Both feedback items
-(`feedback-install-manifest-gap-b773.md` F001/F002) now fully delivered.
-Prior update, same day, `/ardd-plan` — drafted, approved, and tasked
-`plan-install-manifest-gap-fix-2026-07-15-20fb.md` (solo mode, no branch
-gate, on `main`). Consumed both open items of
-`feedback-install-manifest-gap-b773.md` (F001 bug, F002 UX — both
-incorporated; file now `planned`). No unsurfaced defects, no feature
-slugs targeted, no artifact changes (pure packaging-script fix). 5 tasks
-across 2 phases in `tasks-install-manifest-gap-fix-6cfb.md` (`ready`):
-Phase 1 (T001, test-first; T002) adds the missing `cp`/`chmod +x` line
-for `scripts/feature-list.sh` to `install.sh`, proven by extending
-`test-install-worktreeinclude.sh`'s existing Case 1b
-installed-and-executable pattern. Phase 2 (T003, test-first; T004;
-T005) adds `scripts/test-install-manifest-complete.sh` — a new
-packaging-manifest regression check diffing skill-referenced/
-`chmod`-listed scripts against `install.sh`'s `cp` manifest, so this
-class of gap fails CI mechanically going forward — plus its CI job.
-Open question: whether the same manifest-completeness check should
-extend to `templates/*.md`/artifact-template copies too; left out of
-scope for now (the `<name>.sh` grep pattern doesn't generalize cleanly
-to those paths without more design). Prior update, same day,
-`/ardd-feedback` — logged
-`feedback-install-manifest-gap-b773.md`, 2 items (1 bug, 1 UX) discovered
-by manual inspection of an installed project: `scripts/feature-list.sh`
-(added for `list-mode-for-plan-and-impleme`/beta.8 and extended for
-`epics-grouping-in-feature-regi`/beta.9) has no `cp` line in
-`install.sh`'s explicit per-file `ardd-scripts` manifest — confirmed via
-`grep` and direct inspection of `install.sh:169-184` — so it never
-reaches an installed target. Impact: `/ardd-plan --list`, `--epic`
-filtering, and any other `feature-list.sh` consumer are dead outside
-this self-hosted repo; the rest of epic grouping (the `epic:` field,
-`lint-project.sh` validation, `/ardd-status`'s by-epic breakdown,
-`/ardd-tracker`'s milestone assignment) is unaffected since those live
-in already-installed files. F002 flags the missing packaging-manifest
-test that would catch this class of gap mechanically. Not yet consumed
-by a plan. Prior update, same day, `/ardd-implement` — delegated worktree run completed
-and merged all 6 tasks of `tasks-epics-grouping-in-feature-regi-42f2.md`
-(now `completed`): T001/T002 (test-first) added `epic` emptiness
-validation to `lint-project.sh`'s feature-register loop (free-text
-pattern, not the numeric `gh_issue` check), plus fixtures on both sides
-(bad-project empty-`epic` case, good-project non-empty `epic:
-platform-widgets`). T003/T004 (test-first) extended `feature-list.sh`:
-rewrote its arg parsing into a loop to support `--epic <slug>` alongside
-`--status`/`--all`, and added `epic` as a fifth tab-separated output
-column. T005 added an optional "by epic" breakdown to `/ardd-status`'s
-Feature Backlog section (omitted when no feature has `epic` set — the
-case right now, so it stays invisible in this very report). T006 added
-one-directional GitHub-milestone push/assignment to `/ardd-tracker`
-(idempotent creation, register → tracker only, pull never reads a
-milestone back). T005/T006 had no test tasks (Principle V
-documentation-only exception, confirmed before relying on it). Merged
-fast-forward (`6b269e3..1750e8c`, no signing needed) and the worktree
-reaped. Feature `epics-grouping-in-feature-regi` flipped `tasked` →
-`implemented` (rode the branch). Prior update, same day, `/ardd-plan
-epics-grouping-in-feature-regi` —
-drafted, approved, and tasked
-`plan-epics-grouping-in-feature-regi-2026-07-15-d215.md` (solo mode, no
-branch gate, on `main`). No open feedback or unsurfaced defects to
-consume this run. Artifact change: `constitution.md` amended to add
-`epic` to the feature register's documented optional fields (MINOR,
-v1.10.0 → v1.11.0, SIR written) — applied ahead of drafting, on
-confirmation, since the feature's own scope requires it (a documented
-schema change per this repo's semver policy). 6 tasks across 4 phases in
-`tasks-epics-grouping-in-feature-regi-42f2.md` (`ready`): Phase 1
-(T001–T002, test-first) adds `epic` emptiness validation to
-`lint-project.sh` plus fixtures. Phase 2 (T003–T004, test-first) extends
-`feature-list.sh` with an `epic` output column and `--epic <slug>`
-filter. Phase 3 (T005) adds an optional "by epic" breakdown to
-`/ardd-status`'s Feature Backlog section, omitted when no feature has
-`epic` set. Phase 4 (T006) adds one-directional GitHub-milestone mapping
-to `/ardd-tracker`'s push phase — register owns `epic`, pull never reads
-a milestone back. T005/T006 have no test tasks (Principle V
-documentation-only exception; T006 also matches the standing
-mechanization non-goal that `ardd-tracker`'s `gh` glue stays
-judgment-heavy skill prose). Feature `epics-grouping-in-feature-regi`
-flipped `backlogged` → `planned` → `tasked`. Deliberately out of scope:
-no `epic` files with their own lifecycle, no `/ardd-plan` interactive
-epic picker (multi-slug `/ardd-plan <slug> <slug> ...` already covers
-it), no milestone-clearing on `epic` removal (open question). Prior
-update, same day, `/ardd-implement` — delegated worktree run completed
-and merged all 6 tasks of `tasks-list-mode-for-plan-and-impleme-2bf9.md`
-(now `completed`): T001/T002 (test-first) added `scripts/feature-list.sh`
-+ `scripts/test-feature-list.sh` — mirrors `tasks-list.sh`'s glob/
-frontmatter-parse pattern, filters the feature register by status
-(default `backlogged`; `--status <list>`; `--all`), confirmed red state
-before implementing (the pre-commit hook's `scripts/test-*.sh` glob meant
-T001+T002 landed in one commit — noted in the message rather than forcing
-an artificial split). T003 added a matching CI job to `lint.yml`. T004/T005
-added `--list` usage forms to `/ardd-plan` (shells to `feature-list.sh`)
-and `/ardd-implement` (shells to `tasks-list.sh`, filtered to
-`ready`/`in-progress`) — both stop before their skill's first step, no
-writes. T006 documented both in their reference pages. Merged
-fast-forward (`8ae0449..f51eff8`, no signing needed) and the worktree
-reaped. Feature `list-mode-for-plan-and-impleme` flipped `tasked` →
-`implemented` (rode the branch). Prior update, same day, `/ardd-plan
-list-mode-for-plan-and-impleme` —
-drafted, approved, and tasked
-`plan-list-mode-for-plan-and-impleme-2026-07-15-a2c2.md` (solo mode, no
-branch gate, on `main`). No open feedback or unsurfaced defects to
-consume this run. No artifact changes — skill-prose plus one new
-deterministic script; no new principle, data-model concept, or
-production shortcut. Used the new plan-approval browser-preview offer
-(published as an Artifact) for the first time since it shipped. 6 tasks
-across 2 phases in `tasks-list-mode-for-plan-and-impleme-2bf9.md`
-(`ready`): Phase 1 (T001–T003) adds `scripts/feature-list.sh` —
-test-first, mirroring `tasks-list.sh`'s glob/frontmatter-parse pattern,
-enumerating the feature register with a `--status`/`--all` filter
-(default `backlogged`) — plus a CI job. Phase 2 (T004–T006) adds
-`--list` usage forms to `/ardd-plan` (shells to `feature-list.sh`) and
-`/ardd-implement` (shells to the existing `tasks-list.sh`, filtered to
-`ready`/`in-progress` in prose), plus matching reference-doc updates.
-No test tasks for Phase 2 (Principle V documentation-only exception).
-Feature `list-mode-for-plan-and-impleme` flipped `backlogged` →
-`planned` → `tasked`. Open questions: whether `feature-list.sh`'s
-description column needs truncation, and whether `--list`'s
-ready/in-progress filter should cross-reference in-flight worktrees —
-both left to implementation. Separately, a background `fable` subagent
-(dispatched outside this skill's flow, on user request) designed a
-repeatable prerelease-testing exercise from `dev-notes/prerelease-testing-context.md`:
-new `tests/prerelease/` (README runbook, `GUARDRAILS.md`, 7 scenario
-briefs `S1`–`S7`) and a new source-side-only skill,
-`.claude/skills/prerelease-sweep/SKILL.md` (`/prerelease-sweep smoke |
-full | S<n> ...`) — none of this committed yet, sitting untracked for
-review. Prior update, same day, `/ardd-update` — self-hosted (this repo is its own
-ArDD source), so `/ardd-update`'s reinstall step was run directly against
-this checkout to refresh `.claude/skills/` from the current tree (includes
-the v1.10.0 constitution amendment and the codex-second-harness plan from
-the prior step) rather than pulling from elsewhere. All 15 skills, 3
-reference dirs, and all scripts reinstalled clean; all 8 migrations
-already applied (none pending); `.worktreeinclude` and
-`.project/.gitattributes` already correct. `.project/ardd-version.md`
-re-stamped to commit `a656512`. `next_step_prompt`/`delegation`/
-`merge_policy` frontmatter fields already present — no backfill asked.
-Nothing pulled or pushed; no dev-mode source involved (self-hosted, not
-a separate checkout). Prior update, same day, `/ardd-plan
-codex-second-harness-support` — drafted
-`plan-codex-second-harness-support-2026-07-15-f837.md` (solo mode, no branch
-gate, on `main`), then **stopped at the approval checkpoint** on user
-request — the plan stays `status: draft`, not yet tasked. Before drafting,
-applied the feature's own required constitution amendment: `constitution.md`
-bumped 1.9.0 → 1.10.0 (MINOR, SIR written) with a new "Multi-harness
-install" subsection under Project Scope & Intent, documenting
-`install.sh --harness codex`'s five install-time substitutions
-(`AskUserQuestion` → plain-text prompts; `Agent` worktree
-delegation/fan-out → dropped, inline-only; `.worktreeinclude` → not carried
-over; `next_step_prompt` → plain-text offer; and a newly-identified fifth
-substitution the original backlog entry missed, `/ardd-X` → `$ardd-X`
-invocation-sigil rewrite) — this both folds in the de-risking spike's
-findings and formally executes the named scope reversal ("a Claude Code
-skill pack" → primarily Claude Code, also Codex) the feature record flagged
-as required. The plan's 5 phases: Phase 1 is a blocking live
-skill-to-skill-chaining smoke test (the one gate the spike's docs-only
-research couldn't close — a hostile result there aborts Phases 2–5 in favor
-of don't-do-it); Phase 2 adds the `--harness` flag and one substitution
-transformer to `install.sh`; Phase 3 implements and tests each of the five
-substitutions; Phase 4 updates docs (`docs/concepts.md`'s now-stale
-"currently Claude Code-specific" line, README, USAGE.md); Phase 5 is a full
-regression pass. No feedback consumed (none open) and no unsurfaced
-defects. Feature `codex-second-harness-support` stays `backlogged` — the
-`backlogged` → `planned` flip only happens at plan approval (step 11),
-which this run didn't reach. Resume tasking later with `/ardd-plan --from
-plan-codex-second-harness-support-2026-07-15-f837.md`. Prior update, same
-day, `/ardd-implement` — delegated worktree run completed
-and merged all 4 tasks of `tasks-update-channel-switch-flags-c066.md` (now
-`completed`): T001 added `--local`/`--beta`/`--stable` flags to
-`/ardd-update` — `--stable`/`--beta` call `source-resolve.sh --channel`
-directly on the owned checkout and set `ARDD_CHANNEL` for reinstall;
-`--local` resolves a dev-mode checkout (recorded `Source-Path` if already
-`channel=dev`, else prompts for one) and reinstalls from it without
-`ARDD_CHANNEL`; combining flags is a usage error. T002 updated
-`docs/reference/skills/ardd-update.md` to match. T003 added a one-time
-preliminary browser-preview offer to `/ardd-plan` step 10, ahead of the
-existing Approve/Revise/Stop question — publishes `plan.md` via the
-`Artifact` tool and displays the URL, re-firing on each Revise loop back
-to step 10. T004 updated `docs/reference/skills/ardd-plan.md` to match.
-No test tasks (Principle V documentation-only exception, confirmed
-present before relying on it). Merged with a real merge commit (`main`
-had independently advanced with the codex-second-harness research/backlog
-work while this ran — `2749ce1..<merge>`, signed with the on-disk Claude
-key since 1Password was unavailable) and the worktree reaped. Features
-`update-channel-switch-flags` and `plan-approval-browser-preview` flipped
-`tasked` → `implemented` (rode the branch). Prior update, same day,
-`/ardd-backlog` — logged
-`codex-second-harness-support`: single-source Codex CLI support via
-`install.sh --harness codex` installing the same `SKILL.md` files to
-`.agents/skills/` with the ~4 Claude-specific clauses substituted at
-install time (`AskUserQuestion`, `Agent` worktree delegation,
-`.worktreeinclude`, `next_step_prompt`); Codex v1 deliberately
-degraded to inline-only implementation, plain-text prompts, no lint
-hook; no parallel prose tree, no adapter build system. Spec = the
-accepted `/ardd-research` recommendation
-(`research-codex-cli-second-harness-2026-07-15-2d3d.md`). First step is
-a de-risking spike on the go/no-go gates (exact-name skill invocation,
-reliable skill-to-skill chaining) before any install work; requires a
-MINOR constitution amendment (named scope reversal); fallback if the
-spike fails is don't-do-it.). Prior update, same day, `/ardd-implement`
-— delegated worktree run completed
-and merged all 6 tasks of `tasks-discovery-to-work-eager-captur-2b57.md`
-(now `completed`): the artifact→register bridge shipped. T001/T002 added
-terminal "capture newly documented capabilities" steps to `/ardd-init`
-(both paths) and `/ardd-refine` (delta-scoped — the pivot case); T003
-gave `/ardd-status` a detection-only "Documented but Untracked"
-report/STATUS.md section; T004 added `/ardd-backlog --from-artifacts`
-(stable-artifact walk, passage-grounded candidates, one batched
-confirmation, existing `feature-create` path); T005 added `/ardd-defects`'
-routing note (never-built documented scope → backlog, not DEFECTS.md);
-T006 updated the five reference pages (USAGE/guides deliberately
-untouched — no existing sentence became inaccurate). Merged to `main`
-(true merge `da84e79`→signed `2d792da`, `merge_policy: auto`, no
-conflicts — main had advanced with the parallel plan run) and the
-worktree reaped. Features `discovery-to-work-eager-captur` and
-`backlog-sweep-reconcile-from-a` flipped `tasked` → `implemented` (the
-flips rode the branch). Separately, a `/ardd-research` pass vetted Codex
-CLI as a second harness
-(`research-codex-cli-second-harness-2026-07-15-2d3d.md`, uncommitted):
-verdict M-effort single-source port (Codex now reads the same SKILL.md
-format from `.agents/skills/`), inline-only v1, de-risking spike first,
-MINOR constitution amendment required — awaiting user review before any
-backlog entry. Prior update, same day, `/ardd-plan update-channel-switch-flags
-plan-approval-browser-preview` — drafted, approved, and tasked
-`plan-update-channel-switch-flags-2026-07-15-f22c.md` (solo mode, no
-branch gate, on `main`, while a sibling worktree independently worked
-`tasks-discovery-to-work-eager-captur-2b57.md`). No open feedback or
-unsurfaced defects to consume this run. No artifact changes — both
-targeted features are skill-prose-only: neither introduces a new
-principle, data-model concept, or production shortcut, and
-`source-resolve.sh --channel stable|beta` and its dev-mode detection
-already existed, so no script changes either. 4 tasks across 2 phases in
-`tasks-update-channel-switch-flags-c066.md` (`ready`): Phase 1 (T001,
-T002) rewrites `/ardd-update`'s Usage/step 1 to add `--local`, `--beta`,
-`--stable` flags — `--stable`/`--beta` call `source-resolve.sh --channel`
-directly and set `ARDD_CHANNEL` for reinstall; `--local` resolves the
-recorded-or-prompted dev-mode checkout and reinstalls from it without
-setting `ARDD_CHANNEL` — plus the matching reference doc. Phase 2 (T003,
-T004) adds a one-time preliminary browser-preview offer to `/ardd-plan`
-step 10, ahead of the existing Approve/Revise/Stop question: publish
-`plan.md` via the `Artifact` tool, open it, display the URL, then
-proceed to the unchanged three-way choice — re-fires on each Revise loop
-back to step 10 — plus its reference doc. No test tasks (Principle V's
-documentation-only exception). Both features flipped `backlogged` →
-`planned` → `tasked`. Open questions: whether `--local`'s dev-checkout
-path should be remembered across runs, and whether the plan-preview
-artifact needs a stable favicon/title scheme — both left to
-implementation. Prior update, same day, `/ardd-backlog` — logged
-`update-channel-switch-flags`: `/ardd-update` gains `--local`, `--beta`,
-and `--stable` flags to switch the install to the named channel's latest,
-overriding the recorded `Channel:` for the run and re-recording it going
-forward (today the channel is fixed at install-record time; switching
-means hand-editing `ardd-version.md` or reinstalling; `--local` =
-dev-mode against a live checkout, interacting with `source-resolve.sh`'s
-per-channel selection and the dev-mode warn-and-ask path). Prior update,
-same day, `/ardd-plan discovery-to-work-eager-captur
-backlog-sweep-reconcile-from-a` — drafted, approved, and tasked
-`plan-discovery-to-work-eager-captur-2026-07-15-156b.md` (solo mode, no
-branch gate, on `main`): the artifact→register bridge. Consumed all 3
-items of `feedback-artifact-register-bridge-116a.md` (now `planned`). No
-artifact changes — all skill-prose/docs work; no new scripts (the
-capability-vs-design-note judgment stays LLM work per the mechanization
-non-goals) and no schema changes. 6 tasks across 3 phases in
-`tasks-discovery-to-work-eager-captur-2b57.md` (`ready`): Phase 1 —
-terminal capture steps in `/ardd-init` (T001) and `/ardd-refine`
-(T002, delta-scoped for the pivot case); Phase 2 — `/ardd-status`
-advisory "Documented but untracked" section (T003, detection only),
-`/ardd-backlog --from-artifacts` proposal/write mode (T004), and an
-`/ardd-defects` routing note for greenfield unbuilt scope (T005); Phase
-3 — reference docs + lint-docs (T006). Both features flipped
-`backlogged` → `planned` → `tasked`. Open questions: detection noise
-control (tune from atelier, the first intended consumer) and whether
-init's code-based extraction and the new artifact-based capture stay
-distinct prose. Prior update, same day, `/ardd-backlog` — logged
-`plan-time-defrag-slate-analysi`: advisory, recomputed-at-plan-time
-footprint analysis over open backlog items proposing session-optimized
-slates (overlap bundles implemented serially in one plan; pairwise-disjoint
-sets as separate tasks files for worktree fan-out). Its spec is the
-now-completed sync-tab-scroll prototype research report
-(`research-backlog-defrag-slate-analysis-2026-07-15-627c.md` in that
-repo, uncommitted): the cross-item premise degenerated there (N=1 open
-backlog) but the method worked one level down inside
-`phase-2-in-app-authoring`, surfacing the key requirements — dependency
-ordering as a third axis beyond file overlap, graded footprint confidence
-(speculative items never in parallel sets), register-direct status reads,
-and sensible N=0/N=1 handling. Suggested first step recorded in the
-entry: a second research pass against a genuinely large backlog (e.g.
-atelier after `backlog-sweep-reconcile-from-a` backfills its limbo scope)
-before codifying a `/ardd-plan` slate mode. Prior update, same day,
-`/ardd-backlog` — logged
-`plan-approval-browser-preview`, a request for `/ardd-plan`'s plan-approval
-checkpoint to offer rendering `plan.md` as an artifact and opening it in the
-browser (plus displaying the URL), as an alternative to reading raw
-markdown in the terminal. Prior update, same day, `/ardd-feedback` — logged
-`feedback-artifact-register-bridge-116a.md`, 3 UX items documenting the
-missing artifact→register bridge observed across consumer projects:
-discovery limbo (atelier — post-`/ardd-init` foundational scope stranded
-in artifacts with no register entry), pivot limbo (sync-tab-scroll — an
-`/ardd-refine` pivot's new-capability delta had to be hand-backlogged),
-and `/ardd-defects`' drift framing being wrong for greenfield unbuilt
-scope. Three new-capability items were re-filed to the register in the
-same pass (user pre-approved the batch): `discovery-to-work-eager-captur`
-(init/refine terminal step offering to backlog newly-documented
-capabilities), `backlog-sweep-reconcile-from-a` (status advisory
-detection + `/ardd-backlog --from-artifacts` for retroactive limbo), and
-`epics-grouping-in-feature-regi` (optional `epic:` frontmatter grouping
-for release-cadence bundling). A related fourth idea — computed "defrag"
-footprint analysis for plan-time bundling/parallelization — is
-deliberately NOT backlogged yet: a prototype `/ardd-research` run is in
-flight in the sync-tab-scroll consumer repo, and its report will become
-that item's spec. Prior update, same day, `/ardd-implement` — delegated worktree run completed
-and merged all 13 tasks of `tasks-v1-0-0-pre-cut-testing-finding-bf61.md`
-(now `completed`). T001 (test-first): `worktree-align.sh` now refuses
-`aligned=false reason=not-a-worktree` when run from the primary checkout
-rather than a real linked worktree — the fix for the nested-delegation
-collapse S6 found. T007 (test-first): `lint-project.sh`'s Sync Impact
-Report parser now accepts `→`, `->`, and `-->` equivalently, fixing the
-misleading-error bug. T002–T004 strengthened `/ardd-init`'s
-existing-codebase entity cross-checking, added scale-sensitive
-constitution suggestions, and nudged `/ardd-defects` after brownfield
-init. T005–T006 made `install.sh`'s gitignore suggestion a bounded
-`ACTION NEEDED` block and gave `source-resolve.sh` an opt-in, age-gated
-fetch-skip diagnostic (`note=fetch-skipped-fresh-cache`), reusing the
-existing `update_check_max_age_days` field rather than inventing a new
-one. T008–T009 added a `DEFECTS.md` staleness caveat and fixed the
-`ardd-update-check.sh`/`ardd-status` field-name doc mismatch. T010–T013
-were skill-prose polish (`gh pr create` failure path, create-vs-extend
-task phrasing, `/ardd-diagram`'s silent-README note, and the
-`workflow_mode` inline-vs-stamp documentation clarity). Merged to `main`
-(fast-forward, `merge_policy: auto`, `9f111a9..7b65228`) and the worktree
-reaped. No features bound to this plan, so no feature flip. Both feedback
-files (`findings-0344` and `redrive-695b`) fully delivered — all 13 items
-from the v1.0.0 pre-cut testing pass are now fixed. Prior update, same
-day, `/ardd-plan --from` — folded the redrive batch's 7
-items into the existing `plan-v1-0-0-pre-cut-testing-finding-2026-07-15-b89e.md`
-(added Phase 4: deterministic-script fixes — `lint-project.sh`'s Sync
-Impact Report arrow parsing rejecting ASCII `->`, a `DEFECTS.md`
-staleness caveat, and the `ardd-update-check.sh`/SKILL.md field-name
-mismatch; Phase 5: skill-prose fixes — `/ardd-implement`'s undocumented
-`gh pr create` failure path, `/ardd-plan`'s create-vs-extend task
-phrasing, `/ardd-diagram`'s silent README creation, and the
-workflow-field stamped-vs-inline documentation gap) rather than drafting
-a second plan, since both feedback batches came from the same testing
-effort and touch overlapping files. The old 6-task
-`tasks-v1-0-0-pre-cut-testing-finding-32c3.md` (untouched, 0/6) was
-abandoned and replaced with a fresh combined 13-task
-`tasks-v1-0-0-pre-cut-testing-finding-bf61.md` (`ready`) spanning all 5
-phases — worktree-align.sh's distinct-worktree check (T001) remains the
-highest-value fix. Both feedback files now `planned`. Prior update, same
-day, `/ardd-feedback` — logged
-`feedback-v1-0-0-pre-cut-testing-redrive-findings-695b.md`, 7 items (2
-bugs, 5 UX) from a redrive of the 3 v1.0.0 pre-cut scenarios that lost
-their reports to the earlier spend-limit outage (collaborative-mode
-lifecycle, solo inline core loop, peripheral-skills sweep) — this time
-briefed to write reports progressively, and all 3 survived intact. All 7
-scenarios from the original test plan now have genuine, complete
-findings. Highlights: `lint-project.sh`'s Sync Impact Report check
-silently mis-parses an ASCII `->` where it expects the literal Unicode
-`→`, with a misleading error pointing the user elsewhere (F001);
-`/ardd-defects`' `DEFECTS.md` can go stale within the same day with
-nothing in the report signaling it (a claim about `Entity` lacking
-`long_form_keys` was true when checked but false an hour later — F002);
-`/ardd-implement`'s collaborative-mode step has no documented behavior
-for a failed `gh pr create` (F003); `/ardd-plan`'s task phrasing assumes
-existing code even for a project's very first feature (F004); a
-`ardd-update-check.sh`/SKILL.md field-name mismatch (`latest-release=` vs
-documented `source-tip=`, F005); `/ardd-diagram` silently creates a new
-README when none exists (F006); and SKILL.md phrasing doesn't make clear
-that `workflow_mode` is written inline (not `stamp`-ed like the other
-three workflow fields) (F007). On the positive side: S4 confirmed the
-collaborative branch gate and `merge_policy` exclusion work exactly as
-documented; S5's full backlog→plan→implement→status loop built and ran
-real working code cleanly end to end; S7's `/ardd-audit` produced
-genuinely incisive findings (not boilerplate), and `/ardd-lint`,
-`/ardd-refine`, `/ardd-diagram`, and `/ardd-tracker`'s graceful
-`gh`-unavailable degradation all worked as documented. Not yet consumed
-by a plan. Prior update, same day, `/ardd-plan` — drafted, approved, and tasked
-`plan-v1-0-0-pre-cut-testing-finding-2026-07-15-b89e.md`, consuming all 6
-items from `feedback-v1-0-0-pre-cut-testing-findings-0344.md` (now
-`planned`). Notably declined the literal ask in F003 (auto-editing a
-target's `.gitignore` the way `.worktreeinclude` self-applies) as
-conflicting with a standing, deliberate ceiling decision — scoped the fix
-to visibility instead (a distinct, hard-to-miss suggestion block, still
-suggestion-only). No artifact changes: all 6 fixes are skill-prose or
-script behavior. 6 tasks generated across 3 phases in
-`tasks-v1-0-0-pre-cut-testing-finding-32c3.md` (`ready`): T001 (test-first)
-adds a positive distinct-worktree check to `worktree-align.sh` so it fails
-loud instead of silently collapsing onto the primary checkout (F001, the
-highest-value fix — a real pre-1.0 gap in the delegation machinery);
-T002–T004 (parallel) strengthen `/ardd-init`'s existing-codebase entity
-survey, add scale-sensitivity to its constitution-suggestion catalog, and
-nudge `/ardd-defects` as a same-session follow-up on the brownfield path
-(F002/F004/F005); T005–T006 (parallel) improve `install.sh`'s gitignore
-suggestion visibility and add resolution diagnostics to
-`source-resolve.sh`/`/ardd-update` (F003/F006). No feature slugs bound.
-Prior update, same day, `/ardd-feedback` — logged
-`feedback-v1-0-0-pre-cut-testing-findings-0344.md`, 6 items (2 bugs, 4 UX)
-from a pre-v1.0.0 dry-run testing pass: 7 parallel sandboxed scenarios
-covering new.sh acquisition, brownfield reverse-engineer init, consumer
-upgrade + the new `--reconfigure` flag, collaborative-mode lifecycle,
-solo inline core loop, delegated-worktree execution, and a
-peripheral-skills sweep. An account-wide API spend-limit outage killed 3
-of the 7 subagents mid-run and wiped their scratchpad before reports were
-captured (collaborative mode, solo inline core loop, peripheral-skills
-sweep — all reportedly completed execution cleanly but left no usable
-detail); the other 4 produced real findings. Highlights: `worktree-align.sh`
-has no positive check that it's actually in a distinct worktree — a
-nested-delegation test silently collapsed onto the primary checkout
-instead of failing loud (F001); `/ardd-init`'s existing-codebase survey
-missed an entity lacking a colocated schema during reverse-engineering,
-only caught via an explicit spot-check demand (F002); `install.sh`'s
-`.gitignore` guidance is print-only, unlike `.worktreeinclude`'s
-self-applying handling (F003); the constitution-suggestion catalog can
-feel over-built for trivial projects (F004); `/ardd-defects` right after
-a brownfield init caught real drift `/ardd-init` had just introduced,
-suggesting the docs should nudge that pairing (F005); and
-`/ardd-update`'s real tag-resolution path had unclear same-day-publish
-lag when testing the brand-new `--reconfigure` flag (F006). Not yet
-consumed by a plan. Prior update, same day, `/ardd-implement` — delegated worktree run completed
-and merged all 4 tasks of `tasks-redrive-configuration-choices-29ae.md`
-(now `completed`). **First delegation attempt failed live** — the same
-bug the delegation pre-flight check exists for was reproduced anyway: the
-plan/feature/tasks files this run's `/ardd-plan` wrote were never
-committed before delegating, so the fresh worktree aligned cleanly
-(`aligned=true`) but couldn't find `tasks-redrive-configuration-choices-29ae.md`
-and stopped correctly rather than fabricating work. Recovered by removing
-the empty worktree, committing the session's feature/plan/tasks files to
-`main` (`3aa4dfa`), and relaunching. The second run completed: T001
-rewrote `skills/ardd-update/SKILL.md`'s Usage section and step 5 to add a
-`--reconfigure` flag — without it, behavior is unchanged (backfill only
-absent fields); with it, all four workflow fields (`workflow_mode`,
-`next_step_prompt`, `delegation`, and — solo mode only — `merge_policy`)
-are re-asked regardless of presence, showing current values, stamping
-only what changes. T002–T004 updated
-`docs/reference/skills/ardd-update.md`, `docs/reference/configuration.md`,
-and `CLAUDE.md`'s two "asked once" passages to match. Merged to `main`
-(fast-forward, `merge_policy: auto`, `3aa4dfa..95796aa`) and the worktree
-reaped. Feature `redrive-configuration-choices` flipped `tasked` →
-`implemented`. Everything from this session is now on `main`. Prior
-update, same day, `/ardd-plan redrive-configuration-choices` —
-drafted, approved, and tasked `plan-redrive-configuration-choices-2026-07-14-1e00.md`
-(solo mode, no branch gate, on `main`). No artifact changes: constitution.md's
-Governance section already exempts workflow frontmatter fields
-(`workflow_mode`, `next_step_prompt`, `delegation`, `merge_policy`) from
-Sync Impact Report/versioning even when *changed*, not just set. Design
-chosen: extend `/ardd-update` with a new `--reconfigure` flag (over a
-standalone `/ardd-configure` skill or an `/ardd-init --reconfigure` flag)
-that re-asks all four fields regardless of presence, showing current
-values, and stamps only what changes — generalizing step 5's existing
-backfill-only logic rather than duplicating it. 4 tasks generated across 2
-phases in `tasks-redrive-configuration-choices-29ae.md` (`ready`): T001
-rewrites `skills/ardd-update/SKILL.md`'s Usage and step 5; T002–T004
-(parallel, depend on T001) update `docs/reference/skills/ardd-update.md`,
-`docs/reference/configuration.md`, and `CLAUDE.md`'s two "asked once"
-passages to match. No test tasks — all documentation/prose changes, the
-explicit exception under Constitution Principle V. Feature
-`redrive-configuration-choices` flipped `backlogged` → `planned` →
-`tasked`. Prior update, same day, `/ardd-backlog` — logged
-`redrive-configuration-choices`, a request for a way to re-answer/redrive
-one-time configuration choices (e.g. `workflow_mode`, `next_step_prompt`)
-if the user changes their mind after initial setup — currently only
-askable once by `/ardd-init` or `/ardd-update`, or by hand-editing
-constitution frontmatter. Earlier update, same day, `/ardd-implement` — delegated worktree run completed
-and merged all 4 tasks of `tasks-203c-6b16.md` (now `completed`).
-**First delegation attempt failed live**, demonstrating the exact bug
-`feedback-uncommitted-plan-tasks-delegat-a3ff.md` described: the worktree
-subagent aligned cleanly (`aligned=true`) but the plan/tasks files didn't
-exist there because they'd never been committed — it stopped correctly
-per its instructions rather than fabricating work. Recovered by removing
-the empty worktree, committing the session's plan/tasks/feedback/backlog
-files to `main` (commit `f837b7f`), and relaunching. The second run
-completed: T001 added a delegation pre-flight `git status` check to
-`skills/ardd-implement/SKILL.md` (the fix for the very bug just
-reproduced); T002/T003 added the "committing the red state"
-expected-failure-marker note + framework table to
-`templates/constitution-suggestions.md`'s Test-First Development and
-Deterministic Gates entries; T004 mirrored the marker handling in
-`ardd-implement`'s TDD execution step. Merged to `main` (fast-forward,
-`merge_policy: auto`, `f837b7f..e8a6e0c`) and the worktree reaped. No
-features bound to this plan (`features: []`), so no register flip.
-Everything from this session is now on `main`. Prior update, same day,
-`/ardd-plan` — drafted, approved, and tasked
-`plan-203c-2026-07-14-bf43.md` (solo mode, no branch gate, on `main`),
-consuming both open feedback files: `feedback-uncommitted-plan-tasks-delegat-a3ff.md`
-(F001) and `feedback-tdd-xfail-precommit-contradiction-a639.md` (F001,
-F002). Both flipped to `planned`. No feature slugs targeted this run.
-4 tasks generated across 2 independent phases in
-`tasks-203c-6b16.md` (`ready`): T001 adds a delegation pre-flight
-uncommitted-files check to `ardd-implement`'s delegation gate; T002–T004
-add the expected-failure-marker (`xfail`) resolution to
-`templates/constitution-suggestions.md`'s Test-First Development and
-Deterministic Gates entries and to `ardd-implement`'s TDD execution step.
-Deliberately out of scope: this repo's own `.project/artifacts/constitution.md`
-— its test-added-in-the-same-commit paradigm never produces the
-separately-committed red state the contradiction depends on, so nothing
-there needed to change. Prior update, same day, `/ardd-feedback` — logged
-`feedback-tdd-xfail-precommit-contradiction-a639.md`, a design suggestion
-from an Atelier consumer project. Earlier update, same day, `/ardd-backlog` — logged
-`list-mode-for-plan-and-impleme`, a request for `/ardd-plan` and
-`/ardd-implement` to gain a `--list` mode printing eligible
-slugs/tasks-files with basic info, bypassing the interactive pick flow.
-Earlier update, same day, `/ardd-feedback` — logged
-`feedback-uncommitted-plan-tasks-delegat-a3ff.md`, a bug report from
-inspecting a real delegation failure: `/ardd-plan` writes the plan and
-tasks files to disk without committing them, and in solo mode — where
-`/ardd-plan` often runs directly on the default branch — that leaves a
-window where a `status: ready` tasks file is real on disk but absent from
-commit history. `/ardd-implement`'s delegation gate has no pre-flight
-check for this, so a worktree subagent can get `aligned=true` from
-`worktree-align.sh` (which only fast-forwards committed history) and then
-fail cleanly when the tasks file isn't there — after a full agent launch
-round-trip. This is the second observed occurrence of this exact failure
-mode.)
-Keep this current as artifacts are refined and open questions are resolved._
-
-## Artifact Status
-
-| Artifact | Status | Open questions |
-|---|---|---|
-| constitution.md | stable ✅ (v1.12.1; `delegation: eager`, `merge_policy: auto`) | — |
-
-## Open Questions
-
-None in the artifact. One plan-scoped item remains, non-blocking
-(`plan-quickstart-new-project-2026-07-09.md` — optional `gh repo create`).
-
-## Code-vs-Artifact Defects
-
-1 defect — see `DEFECTS.md`, last checked 2026-07-12 (seventh pass). The
-smoke-tier entry (`7efff3a5`) clears when the deliberately-unprovisioned
-`ANTHROPIC_API_KEY` is provisioned; the next `/ardd-defects` pass
-re-derives it. The docs-site work added `mkdocs.yml`, `docs/index.md`,
-and `.github/workflows/docs.yml` (all source-side) — a re-run would
-verify DEFECTS.md against the enlarged doc/workflow surface.
-
-## Feedback
-
-0 open of mine — a file on disk,
-`feedback-codex-scenario-sweep-findings-2eee.md`, belongs to a
-concurrent Codex session and is deliberately left untouched/uncounted,
-see the `_Updated` note above. `feedback-scenario-sweep-2026-07-21-d816-0f18.md`,
-`feedback-new-sh-codex-launch-tty-b951.md`, and
-`feedback-ardd-backlog-codex-syntax-1c0f.md` are all now `planned`,
-consumed by `plan-096b-2026-07-22-479f.md`, `plan-b860-2026-07-22-e518.md`,
-and `plan-35f6-2026-07-22-93b3.md` respectively — see the `_Updated` notes
-above. Delivered today: `feedback-new-sh-codex-undocumented-b8f0.md`
-(1 bug: `docs/install.md` documented `install.sh --harness codex` but
-never mentioned `new.sh --harness claude|codex`'s identical Codex
-support) is now `planned`, consumed by
-`plan-new-sh-codex-docs-2026-07-21-302a.md` — see the `_Updated` note
-above. Delivered earlier today: `feedback-badge-style-variant-followups-dbff.md`
-(2 bugs: F001 the shieldcn badge logo `PLACEHOLDER`; F002 the
-`.worktreeinclude` gap for `.agents/skills/scenario-sweep/`) and
-`feedback-hookspath-side-effect-c707.md` (1 bug: `core.hooksPath` left at
-`/dev/null` by `Agent` worktree creation, alongside the known
-`core.bare=true` side effect) are both now `planned`, consumed by
-`plan-changelog-precommit-2026-07-21-b716.md` — see the `_Updated` note
-above. Delivered earlier today:
-`feedback-badge-style-clash-7db2.md` (1 UX: shieldcn-default badge
-styling direction + interim clash caveat) is now `planned`, consumed by
-`plan-badge-style-variant-option-2026-07-21-982c.md` (caught mid-run
-after `/ardd-plan`'s step 4 open-feedback check initially missed it —
-a strict-equality status match didn't tolerate the file's trailing
-frontmatter comment; surfaced and confirmed with the user post-tasking,
-its interim-caveat ask added as a standalone Phase 0/T010).
-Delivered 2026-07-21 (all 5 consumed by `plan-multi-harness-2026-07-21-76ba.md`, 12 items all accepted, now implemented via d41b193): docs-sweep-ardd-update-harness-28d0 (3 bugs), sweep-cdf0-codex-port-regressi-878c (2 bugs + 3 UX), guardrails-cwd-safety-5f18 (1 bug), plan-preview-prompts-a91e (1 UX), pre-commit-hook-p90-1e7b (1 UX, filed same day from the hook timing investigation). Delivered today: feedback-badge-consumer-feedback-ea66.md (planned, consumed by plan-badge-consumer-fixes-2026-07-21-1c50.md). Delivered 2026-07-20: `feedback-s9-badge-guards-b8b6.md` (3 UX,
-now `planned`, consumed by `plan-badge-guards-2026-07-20-8b60.md`);
-`feedback-badge-renderer-caveat-bf75.md`
-(1 UX) and `feedback-badge-workflow-yaml-ff0c.md` (1 bug + 3 UX) are
-now `planned`, consumed by `plan-badge-slate-2026-07-20-e059.md` — see
-the `_Updated` note above. (`bf75`: `badge.md`'s
-endpoint-vs-dynamic-JSON renderer caveat, drained 2026-07-20; `ff0c`: the
-invalid-YAML badge workflow template with verified fix, a column-0
-heredoc template sweep, branch-filter/`workflow_dispatch` hardening,
-`[skip ci]`), drained 2026-07-20 by `/inbox-drain`'s first live run.)
-Delivered earlier today: `feedback-amend-path-for-skill-written-files-2187.md` (2 UX:
-the amend-path gap for skill-written files; line-citation rot), drained
-2026-07-20 from an out-of-band inbox note, now `planned` — consumed by
-`plan-amend-path-policy-2026-07-20-3c72.md`. Correction: `feedback-repo-critique-6ad1.md`, previously
-reported here as deliberately-left-open, is actually `planned` —
-consumed by `plan-ardd-state-determinism-2026-07-06.md` (2026-07-06).
-Delivered today: `feedback-next-step-prompt-dontask-20da.md`
-(1 UX) and `feedback-source-path-portability-19ce.md` (2 bugs, 1 UX,
-1 Reconsidered) are now `planned`, consumed by
-`plan-dot-project-reviewer-guide-2026-07-20-ee87.md` — see the
-`_Updated` note above. Delivered earlier:
-`feedback-s1-badge-followups-5e84.md` (2 UX items) is now
-`planned`, consumed by `plan-s1-badge-followups-2026-07-20-9667.md`.
-Delivered earlier:
-`feedback-sweep-coverage-expansion-3452.md` (6 UX items),
-now `planned`, consumed by
-`plan-sweep-coverage-expansion-2026-07-20-d493.md` — see the
-`_Updated` note above. Delivered 2026-07-19:
-`feedback-dynamic-badge-discoverability-a123.md` (3 bugs, 2 UX,
-1 Reconsidered) is now `planned`, consumed by
-`plan-dynamic-badge-discoverability-2026-07-19-23cf.md` — see the
-`_Updated` note above. All earlier batches
-delivered — `feedback-prerelease-sweep-2026-07-18-50ea-ad22.md` (the 6
-accepted prerelease-sweep findings) is now `planned`, consumed by
-`plan-prerelease-sweep-fixes-2026-07-18-d341.md` — see the `_Updated`
-note above. All other batches delivered, including `feedback-ardd-status-reference-page-mis-7fa5.md` and
-`feedback-ardd-plan-slate-mode-unrouted-c563.md` (filed by `docs-sweep`'s
-first live dogfood run, consumed by
-`plan-docs-drift-fixes-2026-07-18-2fbb.md`),
-`feedback-ci-migration-tests-unwired-37ee.md` and
-`feedback-prerelease-sweep-scenario-gaps-95f6.md` (both now `planned`,
-consumed by `plan-status-view-mode-2026-07-18-ce1f.md` — see the
-`_Updated` note above); earlier ones include
-`feedback-prerelease-full-sweep-v0-10-1-e5d8.md` (now `planned`,
-consumed by `plan-v0-10-2-fixes-2026-07-17-4465.md` — see the
-`_Updated` note above; `feedback-prerelease-smoke-sweep-849d.md`,
-`feedback-prerelease-full-sweep-62ae.md`,
-`feedback-plan-preview-setting-63b3.md` — all `planned`, consumed by
-`plan-feedback-batch-2026-07-17-e977.md`;
-`delegation-preflight-autocommit-06b1`, `install-manifest-gap-b773`,
-`artifact-register-bridge-116a`, `findings-0344`, `redrive-695b`).
-
-## Audit
-
-`.project/audit.md`: 1 open suggestion (two-channel release paragraph →
-decision-record pointer) + 1 open risk (smoke key unprovisioned, now
-documented as a deliberate standing state). 2 suggestions resolved this
-pass (new.sh tty narrative → decision record, v1.8.1; Governance
-workflow-field exemption, v1.8.2).
-
-## In Flight
-
-- Nothing in flight — `worktree-agent-a0a55c56c6cef36a2` (`096b`,
-  scenario-sweep fixes) merged (`9310fa4`, fast-forward) and was reaped
-  2026-07-22; `worktree-agent-a4c94765b3a3e6564` (`35f6`, Codex
-  `$ardd-*` skill-sigil fix) merged (`9dee9a3`, fast-forward) and was
-  reaped 2026-07-22; `worktree-agent-a2f2806ef246ff447` (`b860`,
-  Codex launch-tty fix) merged (`587d794`, fast-forward) and was reaped
-  2026-07-22; `worktree-agent-a96704d0e644f9288`
-  (`new-sh-codex-docs`) merged (`4f0177c`, fast-forward) and was reaped
-  2026-07-21; `worktree-agent-af5346a43bebe9e32`
-  (`changelog-from-github-releases` + `pre-commit-hook-scoping-expans`)
-  merged (`ff3dabb`) and was reaped earlier the same day;
-  `worktree-agent-a49dbb827b627b722` (`badge-style-variant-option`) merged
-  (`a1de0f2`) and was reaped earlier still; the codex-port worktree was
-  also reaped 2026-07-21 after its branch (`codex/port-foundation`,
-  8a302c8) merged to `main`.
-
-## Recommended Next Step
-
-`/ardd-defects` to refresh the
-2026-07-12 seventh pass against the enlarged surface (now including
-`scripts/release-notes.sh`, the expanded `hooks/pre-commit`, and the
-twice-widened `docs-sweep` scope); the two remaining backlogged features
-(`plan-preview-editor-option`; `codex-second-harness-support` via
-`/ardd-plan --from
-plan-codex-second-harness-support-2026-07-15-f837.md`, gated on its
-Phase 1 live smoke test) wait for later runs; the off-target research
-file (`research-docs-freshness-skill-2026-07-18.md`) is still unconsumed
-on disk; and the one `.project/audit.md` suggestion remains open.
diff --git a/.project/ardd-version.md b/.project/ardd-version.md
index e666c22..e36b8c0 100644
--- a/.project/ardd-version.md
+++ b/.project/ardd-version.md
@@ -1,9 +1,9 @@
 # ArDD Version
 
-_Source: artifact-driven-dev @ 06200bb · Installed/updated 2026-07-24_
+_Source: artifact-driven-dev @ 541b283 · Installed/updated 2026-07-24_
 
 Source-Path: ~/dev/artifact-driven-dev
-Source-Commit: 06200bb4102f0f1a3a8140d5044bb827c3396bed
+Source-Commit: 541b28338c756c55ccb7f9276754204b4f807bee
 Channel: stable
 Harness: claude
 Harnesses: claude
diff --git a/.project/artifacts/constitution.md b/.project/artifacts/constitution.md
index 2f70785..5aaf992 100644
--- a/.project/artifacts/constitution.md
+++ b/.project/artifacts/constitution.md
@@ -267,6 +267,7 @@ next_step_prompt: auto
 delegation: eager
 merge_policy: auto
 workflow_mode: collaborative
+status_history_keep: 5
 ---
 
 # artifact-driven-dev Constitution