Prompt file imported from therealestmatthew/d-system (.claude/commands/idea-triage.md). Fill in {{arguments}} before use. Copyright stays with the author.

Triage open ideas

Drives the idea-triage subagent over every idea currently status: open in _data/ideas.jsonl, records what it finds as a kind: finding annotation, then moves the idea to triaged. This is phase-idea-02 — see PLAN-016 (idea record system) and PLAN-017.04 (annotations) for why the boundaries below are drawn where they are.

This command never moves an idea past triaged. reviewing, promoted and discarded are owner judgement calls (ADR-010, PLAN-016); automating them is the one thing this phase's acceptance explicitly forbids.

1. Find the open ideas

uv run python -c "
from src.db.ideas import load_events, fold
state = fold(load_events())
for idea_id, s in sorted(state.items()):
    if s.get('status') == 'open':
        print(idea_id, '|', s.get('title'))
"

If {{arguments}} names a specific idea id, triage only that one (and only if its current status is open — report and stop if it is not, rather than guessing what the owner meant). Otherwise triage every open idea found above, one at a time.

2. Triage each one

For each open idea, in order:

1. Read its effective title and body from the fold output above — never from _data/ideas.jsonl directly. A raw created event can be stale if the idea was ever amended, since amendments correct title/body without rewriting the original line.

2. Dispatch the idea-triage subagent with that id, title and body. Include this exact line in the dispatch prompt, with <id> already substituted by you, the driver — never left for the subagent to fill in itself:

   When you write your finding, save the text to a file, then run this exact command with
   only the file path changed — do not retype or otherwise reconstruct the idea id yourself:
   uv run python tools/append_idea.py annotate <id> --author agent-idea-triage --kind finding --text "$(cat <your-file>)"
   

   This exists because a prior session had a subagent write one idea's finding text under a different idea's id — it held the right id in its prompt but retyped the wrong one when constructing the command from memory. Handing it the literal command with the id already filled in removes that retyping step at its actual source, rather than trusting the subagent to copy a number correctly across several tool calls.

3. Once the subagent reports the annotate call succeeded, verify it independently before trusting the report — a subagent's self-report is not proof of what it actually wrote (this is exactly how the misdirected write above went undetected until an unrelated audit caught it):

   uv run python -c "
   from src.db.ideas import load_events, fold
   state = fold(load_events())
   anns = state['<id>']['annotations']
   print(anns[-1]['text'][:200] if anns else 'NO ANNOTATION FOUND')
   "

   Read the printed text. If it opens by naming a different six-digit idea id as its own subject (e.g. idea <id>'s newest annotation starts "Idea 000012 proposes..." when <id> is not 000012), the write landed on the wrong idea — stop, do not move status to triaged, and fix it with amend-annotation (see .claude/agents/idea-triage.md's corrected-entry precedent) before continuing to the next idea.

4. Only once verified, move the idea to triaged:

   uv run python tools/append_idea.py status <id> triaged

5. If the subagent found nothing to report, it still writes a finding saying so — every triaged idea has at least one finding annotation, even an empty-handed one. Do not skip the annotation step because there was nothing notable to find.

Do not batch multiple ideas into one subagent call. Each idea gets its own dispatch, so a finding never accidentally conflates two ideas' overlap.

3. Afterwards

Regenerate the markdown view and leave the tree green:

uv run python tools/generate_ideas_md.py
uv run python -m src.governance

Then report to the owner, per idea: its id, title (000046 (idea planner agent), per GOV-006), and the finding in one line — not just "triaged N ideas." A count with no content makes them open the log to find out what was actually found.

Collect every PROPOSED LINK: and PROPOSED PROMOTION: line across the batch into their own lists at the end of your report, separate from the per-idea summary and from each other. These are candidate tools/append_idea.py link / status ... promoted --promoted-to ... calls the subagent found but never executed — the owner reviews and runs (or discards) each one by hand; nothing here writes a link or a promotion automatically. A PROPOSED PROMOTION is the stricter of the two: the subagent only raises one when an idea's actual ask already shipped, not merely when a document touches the same area.