github-issue-tracker: write all identified actions to the task list on check-in to keep the AI focused

[ElliotDrel]

Summary

The skill currently identifies a list of recommended actions in the final report (Step 5a "Recommended Actions" / "Do Today" / "Watch For"), but those actions live only in chat output. As the conversation continues — the user asks a follow-up, takes one of the actions, drafts a comment, etc. — the AI has no persistent record of the original action list. It forgets items, drops follow-ups, and the user has to re-prompt for things the skill already knew about.

This was observed in the 2026-05-03 run: the check-in identified ~7 distinct actions across "Do Today" and "Watch For." The user then started directing actions one at a time ("post the comments," "what about #29934," "do the work and execute"). Without the user explicitly saying "use the task list," several action items would have been lost as context filled up — including discovered subtasks like "verify bun#28026/#28230 merge status" that the AI initially skipped.

Proposed Fix

After Step 5a (compile report), the skill should:

1. Parse the recommended actions out of the report — every "Do Today," "Watch For" with a date, and every concrete next-step from individual issues with next_steps != null.
2. Write each one to the task list via TaskCreate — one task per actionable item, with the source issue noted in the description.
3. Set initial status to pending — these are proposals, not commitments. The user decides what to actually do.
4. Update task status as work progresses — mark in_progress when starting an action, completed when done, deleted if user declines.

The user's role: scan the task list, approve/reject items conversationally. The AI's role: track what's been done vs what's still outstanding without needing the user to re-state requests.

Why This Matters

Without the task list, the failure mode is:

• AI generates a 10-item action list in the report
• User says "do update-tracker leaves orphaned body content when moving issues to Closed #1 and github-issue-tracker: update tracker incrementally as actions happen, not in bulk at the end #3"
• AI does them
• User says "what about that thing with the OAuth?"
• AI has to re-derive what was on the original list, often missing items or guessing
• Context grows, earlier items get evicted, AI now genuinely doesn't remember them

With the task list:

• AI generates the list and immediately writes 10 tasks
• User says "do update-tracker leaves orphaned body content when moving issues to Closed #1 and github-issue-tracker: update tracker incrementally as actions happen, not in bulk at the end #3" → AI marks them in_progress, completes them
• User says "what's left?" → AI calls TaskList, reports the remaining 8 with one-line status each
• Tasks survive context churn, give the user a clean menu

Implementation

In the skill prompt (Step 5 area)

Add to the existing flow:

Step 5b (NEW): Persist actions as tasks
Before presenting the report, parse all "Recommended Actions" and "Watch For" items into TaskCreate calls. One task per action. Description should include: the issue number, what to do, why (linked to the issue's Goal). Mark all as pending.

When the user approves or directs work on a task, mark it in_progress before starting. Mark completed immediately after the action lands (comment posted, PR filed, etc.).

When the user declines an action ("no, skip that one," "don't bother") or says it's no longer relevant, mark the task deleted.

In subagent prompts

Subagents should not create tasks themselves (they're transient). The orchestrator owns task management and creates tasks based on subagent results when new follow-up work is discovered.

Future-run behavior

On the next check-in, the skill should:

1. Call TaskList to see what's still pending from prior runs
2. Include those in the new report under a "Carried Over" section
3. Either resolve them (action was actually taken between runs) or surface them again

Acceptance Criteria

• Skill prompt explicitly directs the AI to call TaskCreate for every actionable item from the report
• Tasks include the source issue reference and a clear single-line subject
• Subagent results that surface new follow-up work get logged as new tasks (not buried in chat)
• On a follow-up message, the AI uses TaskList to know what's outstanding instead of re-deriving from context
• When work completes, the AI marks the task complete before moving on, so progress is visible to the user

Affected Files

• Skill main file — add Step 5b and update Step 5c to reference task IDs
• Subagent prompt templates (in main skill file) — instruct subagents to return structured "follow-up task" suggestions in their summary

Related

• e-stack#2 — batch agents miss PR merge-readiness signals
• e-stack#3 — bulk tracker updates lose intermediate state on interruption

This issue completes the trio. e-stack#2 is about what the skill checks. e-stack#3 is about when state gets persisted to the tracker. This issue (e-stack#4 if it lands as such) is about how the AI keeps itself on task across a long conversation. All three failure modes were observed in the same 2026-05-03 run.