Convex Release Audit

Convex Release Audit

Handle $ARGUMENTS.

Goal: find newer Convex releases, extract work kitcn can actually use, then delegate one concrete implementation slice to $task so it opens the PR.

Goal Template

When this audit needs durable goal state, use the project-owned Convex release audit template instead of the generic task template:

node .agents/skills/autogoal/scripts/create-goal-scratchpad.mjs \
  --template convex-release-audit \
  --title "convex release audit"

This plan owns version discovery, changelog reconciliation, upstream diff evidence, kitcn leverage classification, the selected slice or no-action verdict, and the delegated task prompt. Implementation still goes through task; do not duplicate implementation PR machinery inside the release-audit plan.

Rules

• Use evidence, not vibes. Read both changelog sources and a diff.
• Prefer deleting kitcn glue over adding more glue when upstream fixed the real problem.
• Do not upgrade Convex just because a newer version exists. Ship only a leverageable improvement.
• Keep the PR slice coherent. One release opportunity per PR unless multiple fixes share the same implementation boundary.
• If no actionable opportunity exists, stop with the evidence. Do not open a vanity PR.

1. Establish Current And Target Versions

Find the currently pinned Convex version:

rg -n '"convex":' package.json packages/**/package.json example/package.json

Find the latest published version:

npm view convex version --json

If $ARGUMENTS names a target version, use it as the upper bound. Otherwise use the latest npm version.

Record:

• current pinned version
• target version
• every version in the current-exclusive, target-inclusive range
• exact package files that pin or constrain Convex

2. Read Both Changelogs

Read Ship:

curl -sL https://ship.convex.dev/

Read the upstream npm package changelog through gh, not browser scraping:

gh api \
  -H "Accept: application/vnd.github.raw" \
  repos/get-convex/convex-backend/contents/npm-packages/convex/CHANGELOG.md

Extract only the sections in range. Reconcile disagreements:

• Ship is product-facing signal.
• npm-packages/convex/CHANGELOG.md is package-facing signal.
• If they disagree, keep both facts and investigate in the diff.

3. Read The Upstream Diff With gh

Use a local upstream clone for navigation, creating it only if missing:

test -d ../convex-backend/.git || gh repo clone get-convex/convex-backend ../convex-backend
git -C ../convex-backend fetch origin main --tags

Find refs for the current and target versions. Prefer tags if they exist:

git -C ../convex-backend tag -l "*<version>*" | sort
git -C ../convex-backend log --all --oneline -- npm-packages/convex/package.json

If tags are unclear, inspect version-bump commits in npm-packages/convex/package.json and use the commit before/after each version bump.

Read the compare through gh:

gh api \
  repos/get-convex/convex-backend/compare/<base-ref>...<target-ref> \
  --jq '.files[] | select(.filename | test("npm-packages/convex|cli|agent|dev|auth|codegen|deployment|function|schema")) | {filename,status,patch}'

If the compare is too large, narrow locally after proving the refs:

git -C ../convex-backend diff <base-ref>..<target-ref> -- \
  npm-packages/convex

4. Search Kitcn For Leverage

Search for local Convex integration points and hacks:

rg -n "CONVEX_AGENT_MODE|local-force-upgrade|skip-push|typecheck disable|codegen disable|convex dev|convex init|convex run|CONVEX_|Convex" \
  packages www .agents docs test

Also search institutional notes before proposing work:

rg -i --files-with-matches "convex|upgrade|agent|cli|dev|bootstrap|verify" docs/solutions

Read relevant hits, especially notes about:

• local backend upgrade prompts
• anonymous or non-interactive Convex setup
• kitcn dev, kitcn verify, kitcn init
• hidden Convex flags or leaked upstream plumbing
• docs/skill sync for Convex setup guidance

5. Classify Opportunities

For each release item, classify it:

• feature: new Convex API, CLI command, runtime behavior, or platform feature kitcn can expose.
• compatibility: required work to keep kitcn working with the new version.
• agentic: upstream change that improves non-interactive, deterministic, or machine-readable flows.
• cleanup: upstream change that lets kitcn delete a workaround, hidden flag, fake prompt handling, fallback path, or doc warning.
• no-op: interesting upstream change with no kitcn action.

For every non-no-op, include:

• changelog evidence
• diff evidence
• kitcn file(s) affected
• expected implementation boundary
• verification command(s)
• confidence

Bias toward agentic and cleanup; kitcn exists to make Convex sharper for humans and agents, not to mirror every upstream bullet.

6. Choose One PR Slice

Pick the highest-leverage slice using this order:

1. compatibility breakage
2. delete dirty hack made obsolete upstream
3. agentic CLI unlock
4. product feature kitcn can expose cleanly
5. docs or skill-only update

If the winning slice touches published package code, the delegated task must update the active changeset and run bun --cwd packages/kitcn build.

If it touches scaffold templates, the delegated task must run bun run fixtures:sync and bun run fixtures:check.

7. Delegate Through task

Load $task with a prompt in this exact shape:

Implement this Convex release opportunity.

Current Convex: <version>
Target Convex: <version>

Opportunity: <one-sentence selected slice>
Class: <feature | compatibility | agentic | cleanup>

Evidence:
- Ship changelog: <short citation or summary>
- Convex changelog: <short citation or summary>
- Upstream diff: <refs and files>
- Kitcn evidence: <local files and docs/solutions notes>

Implementation:
- <specific files or boundaries to inspect first>
- <expected code/doc/test shape>

Acceptance:
- <tests/checks>
- <package build if packages/kitcn changes>
- <fixtures commands if scaffold output changes>
- open the PR after verification

Do not preserve obsolete Convex workarounds if the upstream release removes the
need for them. Hard cut the hack.

Then follow task until the PR exists or a real blocker is proven.

Output

Before delegation, keep the audit terse:

Current: <version>
Target: <version>

| Class | Opportunity | Evidence | Decision |
| --- | --- | --- | --- |
| cleanup | ... | ... | selected |

Delegating to task: <selected slice>

After task finishes, use its final handoff format.