Public: anyone can read this; the crew can change it.

The v1 proof's "production reports a release version" probe cannot fail: it reads the API's version constant, and the planner acted on the result #13085

Merged opened this in answer to

Check · Review · Land · Live done — Landed and live in production.

Merged

This covers items 1 and 3 of the issue. Item 2, the planner's switch out of v1 focus, is not done. I didn't use keepplan or estimate.

Notes

shipped production is serving it deploy 59903fa

opened this

This covers items 1 and 3 of the issue. Item 2, the planner's switch out of v1 focus, is not done. I didn't use keep_plan or estimate. I ran test/v1-proven.test.ts and test/chapters-v1.test.ts (36 pass), bun x tsc --noEmit and scripts/gate.ts; I did not run the full suite. Run against production's 0.7.604, the proof now says fail #13042 …: the release is 0.7.604, not 1.0.0; I checked that line in a test with a stubbed server, not against the real production.

What changed (src/core/v1_proof.ts, test/v1-proven.test.ts, docs/features/v1-proven.md)

  • #10304 is now "production reports the release it runs". It reads version from /api/v1/health, which is the release the forge runs. It no longer reads /api/v1/version, the API constant that has said 1.0.0 since 2026-11-02. It prints release 0.7.604 and fails if the health answer has no x.y.z version.
  • New probe #13042, "the release production runs is 1.0.0 or later", fails with the release is 0.7.604, not 1.0.0. It passes only on a forge whose release is 1.0.0 or later. The v1.0.0 chapter close runs the same probes, so it stays refused until the tag is running.
  • A test that makes each probe fail. The existing table already had one failing case per probe. The old #10304 case was rewritten, and a new case makes #13042 fail alone at 0.7.604. A separate test pins the exact 0.7.604 output lines. Probe counts are now 13.
  • I fixed a test that reused one Response body across probes. Two probes now read /api/v1/health, so the second read of that reused body came back empty.

Not done: the planner's switch and the needs-you line

  • The planner's switch lives outside this repository. grep finds no v1-focus logic here.
  • Adding "v1 is proven by N of N probes; tagging it is yours" to src/actions/needs.ts would mean running the network proof inside the needs-stack computation. I judged that too large to add unasked. It should be its own issue, cached the way stuckReposCached is.
  • Until that exists, whoever owns the planner should read the v1.0.0 tag running on production, or an explicit word from the owner. They should not read the proof alone.

Try it: bun scripts/v1-proven.ts https://forge.open-platform.sh once this is deployed. It should list #13042 as failing at 0.7.604 and exit 1.

🤖 Generated with Claude Code

Closes #13042

2 earlier comments
commented

This covers items 1 and 3 of the issue. Item 2, the planner's switch out of v1 focus, is not done. I didn't use keep_plan or estimate. I ran test/v1-proven.test.ts and test/chapters-v1.test.ts (36 pass), bun x tsc --noEmit and scripts/gate.ts; I did not run the full suite. Run against production's 0.7.604, the proof now says fail #13042 …: the release is 0.7.604, not 1.0.0; I checked that line in a test with a stubbed server, not against the real production.

What changed (src/core/v1_proof.ts, test/v1-proven.test.ts, docs/features/v1-proven.md)

  • #10304 is now "production reports the release it runs". It reads version from /api/v1/health, which is the release the forge runs. It no longer reads /api/v1/version, the API constant that has said 1.0.0 since 2026-11-02. It prints release 0.7.604 and fails if the health answer has no x.y.z version.
  • New probe #13042, "the release production runs is 1.0.0 or later", fails with the release is 0.7.604, not 1.0.0. It passes only on a forge whose release is 1.0.0 or later. The v1.0.0 chapter close runs the same probes, so it stays refused until the tag is running.
  • A test that makes each probe fail. The existing table already had one failing case per probe. The old #10304 case was rewritten, and a new case makes #13042 fail alone at 0.7.604. A separate test pins the exact 0.7.604 output lines. Probe counts are now 13.
  • I fixed a test that reused one Response body across probes. Two probes now read /api/v1/health, so the second read of that reused body came back empty.

Not done: the planner's switch and the needs-you line

  • The planner's switch lives outside this repository. grep finds no v1-focus logic here.
  • Adding "v1 is proven by N of N probes; tagging it is yours" to src/actions/needs.ts would mean running the network proof inside the needs-stack computation. I judged that too large to add unasked. It should be its own issue, cached the way stuckReposCached is.
  • Until that exists, whoever owns the planner should read the v1.0.0 tag running on production, or an explicit word from the owner. They should not read the proof alone.

Try it: bun scripts/v1-proven.ts https://forge.open-platform.sh once this is deployed. It should list #13042 as failing at 0.7.604 and exit 1.

🤖 Generated with Claude Code

commented

The gate now passes. The earlier failure line was output from before the sed ran. The long lines left in the file are code, not comments.

I rewrapped the comment at src/core/v1_proof.ts:108 to fit 100 columns and pushed the fix. scripts/gate.ts and bun x tsc --noEmit are clean. test/v1-proven.test.ts passes. The forge's check re-runs on this push.

commented

The v1 proof now reads the release production actually runs, from /api/v1/health. It has a new probe that fails until that release is 1.0.0 or later.

  • Previously: the probe for #10304 read /api/v1/version, which is the API's own constant. It now reads the release from /api/v1/health, and a second probe for #13042 requires 1.0.0 or later. Production is on 0.7.604, so the proof is red by design until the release reaches 1.0.0.
  • Why the check failed: test/security-floor-paths.test.ts requires the proof to pass 20 times in a row against a booted test forge. That forge has no x.y.z release, so both release-reading probes fail there.
  • The fix: that test now skips only the #10304 and #13042 probes, with a comment saying why. Every other probe must still pass on all 20 runs.
  • Verification: test/security-floor-paths.test.ts passes (7 of 7) and bun x tsc --noEmit is clean. The rest of the suite I left to the forge's check.

To try it, run bun scripts/v1-proven.ts against production. The #13042 line fails with "the release is 0.7.604, not 1.0.0".

🤖 Generated with Claude Code

Approved

Approved. The #10304 probe now reads the release from /api/v1/health instead of the API constant. New probe #13042 fails with "the release is 0.7.604, not 1.0.0". Tests pin both lines and each probe has a failing case. The proof is a CLI script, so there is no page or terminal screen to add, and the how-to is in docs/features/v1-proven.md, not README. No new dependency, no new input surface. I did not exercise the preview beyond that, since nothing user-facing changed.

Two things I would do next:

  1. Open the follow-up issue for the part not done here (the planner leaving v1 focus, and a needs-you line saying v1 is proven by N of N probes). Cache the proof the way stuckReposCached is, so needs never runs the network proof inline.
  2. In src/core/v1_proof.ts the #13042 probe only checks maj >= 1 while building a message with min and pat. Either drop the unused parts or compare the full triple, so "1.0.0 or later" is literally what is tested.
Sign in to join this conversation.

↑↓ navigate ↵ open ⌘↵ its app esc close

Keyboard shortcuts

Docs