improve-convex-plugin

This skill ships the current session transcript to anteater's POST /review endpoint, which runs Claude over it and returns structured findings (ambiguous or wrong instructions, places the agent got stuck or repeated work, tooling failures, plus what worked) targeted at the runbook / bootstrap / skill / component. It is a system-improvement loop, not end-user feature feedback.

Consent — ask once, then remember

Sharing a session is opt-in, so you ask once and never nag again. Run the helper (below). If it prints CONSENT_REQUIRED and exits non-zero, the user has not yet chosen — stop and ask them, verbatim intent:

"To improve Convex's tools, this sends a redacted copy of your session to
the Convex team. Secrets are stripped first. Share it — Always (remember and
never ask again), Just this once, or Never?"

Do NOT decide for them and do NOT send anything until they answer. Then re-run the helper with exactly the flag matching their choice:

• Always → --consent always (persists to ~/.convex/improve-consent; future

sessions send with no prompt)

• Just this once → --consent once (shares only this session; not stored, so a

later session will ask again)

• Never → --consent never (persists; never offered again; re-enable by

deleting that file)

Once consent is always, later sessions skip the question entirely. If the helper instead prints CONSENT=never … sharing is off and exits 0, the user has already opted out — just tell them sharing is off and stop; do not re-ask or pass a --consent flag.

How to run it

The helper is served from the Convex backend (so consent-gate and redaction fixes ship instantly, with no plugin update). Curl it and pipe to bash — it auto-detects Claude vs Codex by finding the freshest transcript, compacts + redacts it, uploads, and polls for the review:

curl -fsSL "https://basic-anteater-667.convex.site/send-transcript" | bash -s -- --idea "<the one-line app idea from this session>"

Append the consent flag once the user has chosen, e.g. … | bash -s -- --idea "…" --consent always.

• Pass --idea so the review can correlate to what was being built (read it

from this session's context; omit if unknown).

• --source claude|codex forces the harness; otherwise it auto-detects.
• The URL above already targets the right anteater; --base https://<anteater>.convex.site

(or QBREVIEWBASE) overrides it if needed.

Reading the output

The script prints markers then the review JSON:

• REVIEW_SOURCE=… session=… — which transcript it found.
• REVIEWSUBMITTED id=… — accepted; REVIEWDONE status=done — findings ready.
• The JSON has summary, outcome, findings[] (each with title, target,

severity, observation, evidence, suggestedFix), and wins[].

After it completes, give the user a short summary of the highest-severity findings and what part of the system each one targets. If it prints REVIEWNOTRANSCRIPT, no Claude/Codex .jsonl was found — tell the user. Never paste raw secrets back; the script already redacts keys/tokens before upload, but keep your summary about the system, not the user's data.