platform-docs has no run-lifecycle reference — blocked on #596, and the pages suspected of drift are already correct #601

Activity

1. serge-ivo

   added a commit that references this issue on Aug 15, 2026

   docs(runs): the run-lifecycle vocabulary is published, and its value …

   ...

   7a19d10
2. serge-ivo commented on Aug 15, 2026

   

   serge-ivo

   on Aug 15, 2026

   ContributorAuthor

   More actions

   Delivered in 7a19d104. CI green, host deploy green, verified live.

   The page

   platform-docs/run-lifecycle.md, nav 10 → 11 both directions. Live at
   https://proagentstore.online/docs/run-lifecycle/ — 200, linked from the docs index, and the
   rendered park table reads one "gives up" (human) and two "resumes" (engine_limit,
   platform_interrupt), which is #596's semantics arriving intact on the public page.

   AC1 — generated, not transcribed

   The two value sets are rendered into delimited regions from RUN_HEALTH_STATES,
   RUN_WAIT_REASONS and PARKS; the gate compares them byte-for-byte. Only the value sets are
   generated — the prose stays hand-written, per the rule wire-surface.mjs states about golden
   files over explanation.

   PARKS.why is rendered too, as a considered trade: it reaches the owner verbatim through
   waitClause, so a paraphrase could disagree with what the platform actually says. The cost is
   that rewording a park regenerates the block.

   Every failure mode was watched, not assumed

   Both directions of drift, driven end-to-end before the guard was trusted:

   Mutation	Result
   5th RUN_HEALTH_STATES member, page untouched	red, names the count
   4th RUN_WAIT_REASONS member with no PARKS row	red, names the reason
   human flipped give_up → resume	red — the #596 regression, which a set-membership check would not catch
   enum reshaped to a bare type union	red as UNREADABLE, not silently empty (G3)
   page edited back to the pre-#588 wording	red, printing the expected text verbatim
   unmutated control	green

   The code-side cases were driven through in-memory mutations of workers/api sources, which
   another agent holds live. Nothing under workers/api/** is touched.

   One defect this found in itself

   The first end-to-end red run printed the ✓ denominator line beside the ✗. A success line next
   to a failure about the same check reads as "mostly fine" — the #604 collapse wearing a summary's
   clothes. The note is now withheld whenever the check failed, pinned by a test.

   AC4 — denominator

   ✓ run lifecycle: 2 generated region(s) in platform-docs/run-lifecycle.md == 4 health state(s) + 3 park reason(s) (2 deadline kind(s)) from 2 code file(s)
   

   Floors are derived, not observed: 3 health states (the pre-ended size; runHealth returns
   four distinct literals) and 2 park reasons (the smallest set carrying both a resume and a
   give-up). Below either, the parser has stopped parsing — and an empty render compares equal to an
   empty region, so without the floor this guard would pass by checking nothing.

   One correction to the brief

   I was pointed at work-report.ts:84-88 as the reference wording for "neither timestamp speaks for
   the engine". Those lines are the ended-is-a-member rationale. The engine caveat is at
   work-report.ts:314-317 ("What it deliberately does NOT claim: that the ENGINE is working…
   The live runState lives behind /capture"). The page cites the latter; both facts are stated,
   under the right citations.

   Verified / not verified

   • Verified: all 14 docs:drift checks, pnpm docs:build (zensical --strict), pnpm typecheck
     (11 projects), every scripts/check-*.mjs gate enumerated from ci.yml (including
     check-surface-lock --require-history and check-mcp-parity), biome, CI green on 7a19d104,
     and the live page + nav.
   • Full suite: measured on a throwaway worktree at origin/main + only these five files —
     all 10 scripts/ test files, 168 tests, green. The main checkout carried failures in
     workers/mcp/src/** from another agent's uncommitted work; those same files pass at clean
     origin/main, so they are neither mine nor main's. No git stash was used at any point.
   • Not run: pnpm test:e2e. Nothing under store/ or agents/coder/web was touched, and the
     only docs path any spec loads is /docs/browser-runtime/ (e2e/console-server.mjs:209).
   • Not observed live: no run is parked right now, so the page's park table has not been read
     beside a real waiting row. Same gap waiting_until means 'expected to resume', and the human handoff's deadline is a give-up — one field, two opposite actions #596 recorded; the next handoff will carry it.
3. serge-ivo

   closed this as completedon Aug 15, 2026