feat(#4976): store mint and inference settings in per-repo config

[fullsend-ai-coder]

Summary

• Add mint_url, inference_provider, inference_project, inference_region, and inference_wif_provider fields to per-repo config schema (ADR 0069 Decision 1)
• Implement accessor methods (ConfigMintURL(), ConfigInferenceProvider(), etc.) on PerRepoConfigReader with layered fallback (overlay → base → code defaults) and corresponding setters on PerRepoConfigWriter
• Wire fullsend github setup to populate mint/inference values in config: flag values go into config.yaml (no preset) or overlay (with preset), while the base layer stays identical to the fetched preset

Design

Schema fields: Flat YAML keys on perRepoConfig (mint_url, inference_provider, inference_project, inference_region, inference_wif_provider). Field names are prefixed to avoid colliding with the org-mode detection keys in IsPerRepoYAML (dispatch, inference).

Accessor pattern: Follows the existing scalar fallback convention (e.g., ConfigRuntime(), ConfigForge()): local value → parent → empty string. Code defaults in perRepoDefaults return "" for all mint/inference fields — callers that need a fallback (e.g., the CLI) use DefaultMintURL.

Installer wiring: When --config (preset) is provided, flag-specified values go into the overlay; unspecified flags inherit from the base layer via the accessor chain. The changedFlags map records which CLI flags were explicitly set.

Backward compatibility: Existing repo variable/secret writes (FULLSEND_MINT_URL, FULLSEND_GCP_REGION, FULLSEND_GCP_PROJECT_ID, FULLSEND_GCP_WIF_PROVIDER) are preserved so existing workflow templates continue to read from vars/secrets when config fields are absent.

Testing

• Config accessor fallback: overlay → base → defaults for all 5 new fields
• Marshal roundtrip: YAML serialize → parse → accessor match
• Empty overlay omits inherited fields (no leak from base to overlay marshal)
• IsPerRepoYAML correctly identifies config with new fields as per-repo (not org)
• buildPresetOverlay correctly writes only flag-changed values
• Validation rejects invalid inference_provider values
• Existing tests continue to pass unchanged

---

Closes #4976

Post-script verification

• Branch is not main/master (agent/4976-mint-inference-config)
• Secret scan passed (gitleaks — fac93d227b4e30eaa553b16d059eac6624ec9fc2..HEAD)
• PR body secret scan passed (gitleaks — no-git)

[fullsend-ai-review]

🤖 Finished Review · ✅ Success · Started 12:13 PM UTC · Completed 12:28 PM UTC
Commit: 9492ea9 · View workflow run →

[codecov]

Codecov Report

❌ Patch coverage is 92.55319% with 7 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
internal/config/interfaces.go	85.71%	5 Missing ⚠️
internal/cli/github.go	94.44%	1 Missing and 1 partial ⚠️

📢 Thoughts on this report? Let us know!

[fullsend-ai-review]

Review

Findings

Medium

• [architectural-coherence] internal/cli/github.go:298 — SetInferenceProvider("vertex") is hard-coded in the non-preset setup path with no corresponding --inference-provider CLI flag. Unlike the other four mint/inference settings, users cannot specify this value via the command line. Functionally correct today (only one valid provider), but creates flag-set asymmetry.
  Remediation: Add --inference-provider flag defaulting to "vertex", and include "inference-provider" in buildPresetOverlay's flagNames.

• [missing-doc] docs/guides/infrastructure/layered-config-reference.md:71 — The per-field merge rules table and code defaults reference table do not document the five new config fields (mint_url, inference_provider, inference_project, inference_region, inference_wif_provider). This is the canonical layered-config reference.
  Remediation: Add entries for all five fields as scalar-override fields with empty-string code defaults.

• [stale-doc] docs/cli/github.md:45 — CLI flags table documents --mint-url, --inference-project, etc. but does not mention that values are now persisted to .fullsend/config.yaml fields.
  Remediation: Note that flag values are written to the per-repo config file during setup.

Low

• [data-exposure] internal/cli/github.go:300 — GCP project ID and WIF provider resource name are now in both .fullsend/config.yaml (plaintext) and GitHub repo secrets. These are resource identifiers (not credentials), and dual storage is likely transitional for backward compatibility.

• [input-validation] internal/config/config.go:773 — Validate() checks inference_provider against ValidProviders() but does not validate mint_url (HTTPS), inference_wif_provider (format), or other new fields when loaded from disk. CLI validates at the flag level, but configs loaded via LoadConfig bypass those checks.

• [design-smell] internal/config/config.go:206 — Comment claims field names are "prefixed to avoid colliding with org-mode detection keys" but names avoid collision by specificity (inference_provider vs inference), not by a shared prefix.

• [incomplete-doc] docs/ADRs/0069-ready-made-configuration-presets.md:71 — Decision 1 doesn't specify the exact YAML field names implementing the decision.

• [incomplete-doc] docs/guides/getting-started/configuring-github.md:62 — Setup examples don't explain where config values are stored post-installation.

• [pattern-inconsistency] internal/config/interfaces.go:375 — Section comment block explaining the fallback pattern is more verbose than equivalent sections for existing scalar getters.

• [comment-style] internal/config/interfaces.go:346 — ADR reference in interface getter doc comments inconsistent with other getters in the file.

• [comment-style] internal/config/defaults.go:40 — Default getter comments more verbose than established concise pattern.

• [incomplete-doc] docs/cli/github.md:56 — --config flag description doesn't explain how flag-specified mint/inference values interact with presets (overlay behavior).

---

Labels: PR modifies per-repo config schema and CLI setup for mint/inference settings

[fullsend-ai-review]

See the review comment for full details.

---

Note: The following inline comments could not be posted on the diff (GitHub returned 422) and are included here instead:

• internal/cli/github.go:298: [medium] architectural-coherence

SetInferenceProvider("vertex") is hard-coded in the non-preset setup path with no corresponding --inference-provider CLI flag. Unlike the other four mint/inference settings, users cannot specify this value via the command line. Functionally correct today (only one valid provider), but creates flag-set asymmetry and prevents users from selecting a different provider if one is added to ValidProviders().

Suggested fix: Add --inference-provider flag defaulting to "vertex", and include "inference-provider" in buildPresetOverlay's flagNames list.

• internal/cli/github.go:300: [low] data-exposure

GCP project ID and WIF provider resource name are now written to both .fullsend/config.yaml (plaintext) and GitHub repo secrets. These are resource identifiers (not credentials), and dual storage is likely transitional for backward compatibility, but the authoritative source should be clarified.

• internal/config/config.go:773: [low] input-validation

Validate() checks inference_provider against ValidProviders() but does not validate mint_url (HTTPS), inference_wif_provider (format), or other new fields when loaded from disk. CLI validates at the flag level, but configs loaded via LoadConfig bypass those checks.

Suggested fix: Consider adding HTTPS validation for mint_url and pattern validation for inference_wif_provider in Validate().

• internal/config/config.go (file-level): Line 206 · [low] design-smell

Comment claims field names are 'prefixed to avoid colliding with org-mode detection keys' but names avoid collision by specificity (inference_provider vs inference), not by a shared prefix.

Suggested fix: Update comment to accurately reflect the naming convention.

• internal/config/interfaces.go (file-level): Line 375 · [low] pattern-inconsistency

Section comment block explaining the fallback pattern is more verbose than equivalent sections for existing scalar getters.

• internal/config/interfaces.go (file-level): Line 346 · [low] comment-style

ADR reference in interface getter doc comments inconsistent with other getters in the file which don't include ADR references.

• internal/config/defaults.go:40: [low] comment-style

Default getter comments more verbose than established concise pattern used by other default methods.