Bug: Blank responses when resuming certification after context limit in new session

[geneticallymodifiedfoodforthought]

Problem

When a learner hits a context limit mid-certification and starts a new conversation to resume, Addie returns blank responses. The session cannot be resumed.

Confirmed case

• Escalation Add unified contacts view with engagement scoring #462 — John Castillo, attempting to resume A3 certification. Steps: hit token limit → start new conversation → say "resume A3 certification" → blank response. Device: MacBook, Chrome.

Expected behavior

Addie should detect the interrupted certification session and resume teaching from the last checkpoint.

Actual behavior

Blank/empty response is returned. The learner is stuck with no path forward.

Steps to reproduce

1. Start a certification module (confirmed on A3)
2. Continue until context/token limit is hit
3. Start a new conversation
4. Say "resume [module] certification"
5. Observe: blank response

Notes

• Distinct from GH Bug: Daily Addie conversation limit not resetting after 24 hours #6048 (daily limit reset bug) — this is about session resumption after a context window limit, not a daily cap
• Checkpoint system exists but may not be surfacing correctly across session boundaries

Related

• Escalation Add unified contacts view with engagement scoring #462 (John Castillo)
• GH Bug: Daily Addie conversation limit not resetting after 24 hours #6048 (daily limit reset — related but separate)

[bokelley]

Thanks for filing!

Triage

Classification: Bug
Bucket(s): addie, training/certification
Status: ready-to-implement

What the experts said:

• prompt-engineer: The web path (addie-chat.ts) already registers CERTIFICATION_TOOLS at line 591 — missing tool registration is not the root cause. Most likely: the stream_error handler (lines 1264–1281) calls res.end() without emitting a text fallback, so the client receives nothing when the model aborts mid-stream. A secondary gap: null workos_user_id at line 539 silently skips buildCertificationContext for some web sessions.
• education-expert: Confirms cert tools are already wired on the web path. The checkpoint system stores enough granularity (concepts covered, learner gaps, demonstrated skills) to resume meaningfully once the stream fix lands. Recommends a follow-on issue: enforce a minimum-one checkpoint within the first 3 turns so context-limit hits don't cold-restart the module.

My take: Silent P0 — user sees nothing, no error, no retry affordance. Note: Escalation #462 (John Castillo) also appears in #6052 (non-member paywall gap); confirm whether his session state affects the cert context lookup before assuming stream_error is the sole trigger.

• Confirm failure mode first: Check server logs for Addie Chat Stream: Stream interrupted mid-reply or addie_empty_response_fallback on Castillo's session — narrows stream_error vs. earlier failure.
• Fix A (addie-chat.ts ~line 1264): Before res.end() in the stream_error handler, emit a text event with EMPTY_RESPONSE_FALLBACK — mirrors the done-path guard already at lines 1291–1303.
• Fix B (addie-chat.ts line 539): If logs show cert context absent, investigate null memberContext?.workos_user?.workos_user_id; add a logger.warn so this failure mode is observable rather than silent.
• No protocol changeset. CERTIFICATION_TOOLS are already registered in addie-chat.ts — no tool registration change needed.
• Follow-on: enforce min-one checkpoint within first 3 turns (separate issue).

---

Triaged by Claude Code. Session: https://claude.ai/code/session_011wYdzP8Ry1R1XWDTbikCDe

---

Generated by Claude Code