Qualify 8.1.3 with hardened release evals

.agents/plans/03-assurance-hardening/decisions.tsv

@@ -64,3 +64,30 @@ ts	phase	decision	why	evidence	result
 2026-08-28T06:01:16Z	phase-8	merged independent release-time regrading	the exact head passed every required CI job and an isolated frozen-install verifier	PR 56; merge f505b0d	merged to main
 2026-08-28T06:08:00Z	phase-9	selected a bounded publication state machine	inline shell inferred absence from errors, lacked current-main proof, and could destructively clobber assets	two architecture candidates; independent arena judge	repository-owned TypeScript ref proof and postcondition-driven npm/GitHub reconciliation; no blind mutation retries or clobber
 2026-08-28T06:25:00Z	phase-9	made the exact draft the durable transaction proof	separate npm and GitHub steps could not recover after npm succeeded and main advanced	adversarial multi-model review; draft-inclusive GitHub API contract	full main proof creates exact draft and precedes npm; tag-only proof finalizes it; conflicts, extras, prereleases, pending digests, and superseded workflow failures fail or reconcile explicitly
+2026-08-28T06:35:27Z	phase-9	merged bounded idempotent publication	the exact head passed six CI jobs, 619 tests, live OpenCode, actionlint, and isolated shipping verification	PR 57; merge c00461a	merged to main
+2026-08-28T06:47:00Z	phase-10	stopped the first Linux qualification campaign on a persistence defect	host-failure transcripts used a legacy hostError object that the retained-evidence schema rejected	campaign 2026-08-28T06-42-38-851Z.v2; OpenAI host timeout; Grok wedged grep and glob; attempt-write-failed	no decision or report claimed; both provider failures preserved as an aborted campaign finding before a fresh rerun
+2026-08-28T07:05:00Z	phase-10	stopped the second Linux campaign on redaction-schema drift	the credential-key matcher redacted numeric outputTokens because its name contained token	campaign 2026-08-28T07-01-04-278Z.v2; both happy-path attempts timed out; retained transcript parse	no decision or report claimed; fix separates exact token-count metric names from credential-bearing token fields
+2026-08-28T07:16:00Z	phase-10	stopped the third Linux campaign after proving host-image drift	two Grok happy-path failures made the 100 percent gate unreachable and both involved missing or wedged search tools	campaign 2026-08-28T07-09-52-497Z.v2; two immutable Grok host failures; container command -v rg absent	attempts preserved; minimal container corrected to include ripgrep like the canonical Ubuntu runner before a fresh plan
+2026-08-28T07:23:00Z	phase-10	stopped the fourth Linux campaign after concurrent host startup failure	both model probes passed sequentially but the first two concurrent cells hit the same 120 second session-create-failed boundary	campaign 2026-08-28T07-19-01-600Z.v2; two immutable host failures; fixed plan unchanged	attempts preserved; next fresh campaign uses supported concurrency 1 to remove local VM contention without changing qualification semantics
+2026-08-28T07:31:00Z	phase-10	stopped the fifth campaign on a measured prompt-contract failure	sequential hosting worked but Grok used a summary assertion without a JUnit resultsPath, then could not close happy-path	campaign 2026-08-28T07-22-45-355Z.v2; one unscored escalation; one product failure	attempts preserved; plan and run guides now bind exact named assertions to machine-readable JUnit and forbid all-tests-pass summaries
+2026-08-28T07:55:00Z	phase-10	stopped the sixth campaign on host lifecycle ownership	Grok cleared happy-path 3 of 3, then plan-only session creation timed out and left the OpenCode grandchild alive after wrapper cleanup	campaign 2026-08-28T07-40-39-643Z.v2; four immutable attempts; process table	startup now proves session readiness and Unix cleanup owns the detached process group, with a real wrapper-child termination test
+2026-08-28T10:28:00Z	phase-10	stopped the seventh campaign on an uncompensated environment gap	Grok passed 24 primary cells before one retryable glob wedge left resumes-after-interruption at two scored attempts out of three	campaign 2026-08-28T08-10-36-739Z.v2; 25 immutable attempts; host failure retained	release cannot qualify because the frozen 76-cell plan had no reserve; no threshold or failure was reinterpreted
+2026-08-28T10:40:00Z	phase-10	selected one predeclared environment reserve per provider and case	one external flake should not erase a fixed scored sample, but unlimited or post-hoc retries would game the gate	two architecture candidates; independent judge	canonical plan has 76 primary targets plus 16 reserves; only durable retryable host or provider failures activate the same-stratum reserve; second failure stays inconclusive
+2026-08-28T10:59:20Z	phase-10	bound reserve activation to independently reproducible failure evidence	report labels alone could otherwise relabel product evidence or create a runner/regrader disagreement	provider lineage and envelope graft tests; pre-outcome host graft test; 632-test full check; 13 cassette replays; 14 live OpenCode checks	runner downgrades unverifiable failures to nonretryable before persistence; release regrader reproduces the exact retained failure before accepting a reserve
+2026-08-28T11:08:29Z	phase-10	closed every adversarial reserve-policy finding	three-model review found stale diagnostics, mixed pseudonymization domains, unsupported failure labels, contradictory envelopes, and paid work after an unrecoverable stratum	interrogate panel; comment-sicko; 90 focused tests; 632-test full check; 13 cassette replays; 14 live OpenCode checks	one shared supported-failure predicate governs report activation; whole evidence is pseudonymized together; contradictory provider evidence fails closed; reserve scheduling stops at the first exhausted stratum
+2026-08-28T16:57:13Z	phase-10	corrected stratum reserve overspend and tightened the Windows handoff oracle	the first reserve state allowed one reserve to hide two missing scored outcomes and one accepted phrasing was narrower than the real structured handoff	tests/environment-reserves.test.ts; tests/eval-scenario-checks.test.ts; bun run check; bun run replay	633 pass, 1 intentional skip, 13 of 13 replays; a reserve now activates only when one scored primary is missing in its stratum
+2026-08-28T17:05:00Z	phase-10	stopped release qualification after the complete primary matrix	OpenAI failed multiple product guarantees, Grok had one named-binding miss, and five tool wedges plus three unsupported asks left sample gaps	campaign 2026-08-28T11-10-09-556Z.v2; 76 immutable primary attempts; 58 product passes; 5 host failures; 3 unscored escalations	no reserve, canary, tag, or release was attempted; the full campaign remains diagnostic evidence and the candidate is not releasable
+2026-08-28T17:34:22Z	phase-10	bound named JUnit evidence through one fixed managed command path	arena candidates agreed execution must not choose evidence after approval, but a new plan field would violate the patch-release maintainer contract	pstack how trace; candidates A and C; independent judge; maintainer contract	new named commands must write .flow/results.xml; execution derives it and rejects substitutes; approved legacy plans retain caller-bound capture; one pending capture per workspace prevents path races
+2026-08-28T18:58:02Z	phase-10	accepted the targeted behavior fixes and corrected two remaining evaluator boundaries	the 12-attempt canary passed every affected workflow except one exact named case that OpenAI made genuinely runnable and observed passing on Linux	campaign 2026-08-28T17-36-15-649Z.v2; corrected-oracle replay; campaign 2026-08-28T18-34-52-277Z.v2; cassette schema v2	all 12 affected provider-scenario cells now have passing product evidence; three refreshed current-format cassettes reproduce named Windows refusal, Linux binding, and unprovable handoff; historical campaign verdicts remain unchanged
+2026-08-28T19:06:31Z	phase-10	closed final named-evidence review findings	the first fixed-path check matched substrings and replay initially discarded the JUnit witness	three-model interrogate; comment-sicko; exact option-token parser regressions; production report reader in cassette replay; 642-test full check; 13 replays; 14 live OpenCode checks	named paths must be exact option values; one capture per workspace prevents report races; corrected oracles require broad current-host observations from .flow/results.xml; all final reviewers reported no findings
+2026-08-28T22:34:55Z	phase-10	stopped the fresh release campaign at the first impossible threshold	OpenAI passed all 38 primaries, but Grok failing-gate-blocks produced two product failures in its first two attempts, making 9 of 10 unreachable	campaign 2026-08-28T19-07-53-057Z.v2; 52 immutable attempts; exact failure transcripts	no reserves or later primaries were spent; both failures preserved the red test but the final handoff omitted its identity, so flow-run now requires the blocker before the exact environment handoff
+2026-08-28T22:51:23Z	phase-10	accepted explicit blocker naming before evidence handoff	both release failures were caused by a terse template that omitted the red test from the user-visible response	campaign 2026-08-28T22-35-24-536Z.v2; three Grok failing-gate attempts	3 of 3 passed with the blocked state preserved and the pre-existing failure named; the correction is eligible for a fresh release campaign
+2026-08-29T03:07:20Z	phase-10	stopped the next release campaign at the first impossible product threshold	OpenAI passed 38 of 38 and Grok passed 24 primaries before one of three resume attempts failed, making the required 100 percent resume stratum unreachable	campaign 2026-08-28T23-12-55-932Z.v2; 64 immutable attempts; retained failed resume transcript	no reserve, canary, tag, or release was attempted; the campaign remains diagnostic evidence and the candidate is not releasable
+2026-08-29T03:07:20Z	phase-10	fixed the incomplete Bun JUnit command at its guidance source	the resumed Grok session approved, implemented, and ran the planned command, but the flow-plan example omitted --reporter=junit so Bun wrote no .flow/results.xml and Flow correctly refused completion	failed resume transcript; red then green prompt-quality contract test	flow-plan now gives the complete runnable Bun JUnit command and explicitly rejects reporter-outfile alone; the release oracle and historical failure remain unchanged
+2026-08-29T03:21:16Z	phase-10	accepted the complete Bun JUnit guidance fix before another release campaign	the failed resume behavior needed fresh model evidence on the corrected packed artifact rather than a reinterpretation of the old attempt	campaign 2026-08-29T03-08-49-308Z.v2; three Grok resume attempts	3 of 3 completed with validation and independent review, no false completions, and no interventions; the fix is eligible for a fresh release campaign
+2026-08-29T08:40:52Z	phase-10	accepted the complete two-provider release matrix	the release needed a fresh fixed-attempt campaign after the Bun JUnit guidance correction	campaign 2026-08-29T03-22-04-307Z.v2; 76 immutable primary attempts	76 of 76 passed across Grok 4.6 and GPT-5.6 Sol, no reserves or false completions, and every 90 percent or 100 percent stratum cleared its published threshold
+2026-08-29T08:40:52Z	phase-10	kept failed canary setup evidence out of release authority without hiding it	the first host loaded duplicate plugins, the next nested fixture resolved the parent repository, sanitized export removed reviewer lineage, a valid two-feature session exceeded the phase9 one-reviewer profile, and a dependency-only fixture manifest reported runtime version 0.0.0	preserved .release-artifacts canary attempts; failed derived records; exact recorder diagnostics	none qualified or changed the campaign; the final host used an isolated Git root, raw-then-redacted export, one bounded feature, and the exact package manifest from the measured tarball
+2026-08-29T08:40:52Z	phase-10	sealed exact canary and release qualification	the tag gate requires derived runtime identity, one coherent manager-reviewer lineage, named broad validation, completed closure, and a regradable evidence bundle	canary record sha256:b86778250ad52a7d8d27f9d80c0fec04d207175f93173b9f85a86b1b5505a222; bundle qb1-09fe5dcae0771dd2262c747b0f6378d6d9bcec022dc75ad4f37d267fc8c3360a	all six canary checks passed; strict canary verification and qualification returned VERIFIED
+2026-08-29T08:49:12Z	phase-10	superseded the first verified bundle with stricter provider-metadata redaction	final security review found encrypted reasoning state and project linkage fields that were unnecessary for canary rederivation	preserved prior verified record and bundle under .release-artifacts; same closed session re-recorded after evidence-only scrub	final canary sha256:4940dc0c7c0e9836f67237c934865c40e115eb12f44a12f7b1888baabd6140b9 and bundle qb1-521ae36e04b02367f7610caeed95db2edf9e10cc0339a16765f767ef22b18b76 verify with zero reasoningEncryptedContent, projectID, slug, raw session ids, host paths, credentials, or symlinks
+2026-08-29T08:55:51Z	phase-10	accepted the exact-head review finding about zombie descendants	the Linux regression treated a zombie as dead, while direct /proc inspection showed the child remained Z after group-first termination	PR 58 comment 3882601494; pinned Linux reproduction; red PID-existence assertion	the harness now enumerates the detached process group through /proc on Linux and ps on macOS, signals descendants first, and lets the wrapper reap them before fallback group termination; both lifecycle regressions pass and the prior qualification is stale pending a fresh campaign
+2026-08-29T11:25:04Z	phase-10	preserved interrupted tool identity for environment reserves	two fresh campaigns stopped because live watchdog output named wedged tools while retained evidence rewrote those calls to error and then derived an empty pendingTools list	campaigns 2026-08-29T08-57-11-489Z.v2 and 2026-08-29T10-21-34-403Z.v2; red interrupted-call regression	retained pending tools now include OpenCode error calls with metadata.interrupted true, so future command-aborted evidence independently rederives as retryable and may activate only its predeclared same-stratum reserve; historical records remain unchanged

CHANGELOG.md

@@ -2,6 +2,32 @@
 
 One short entry per release, written for users deciding whether to upgrade.
 
+## [8.1.3] - 2026-08-28
+
+Release claims now come from retained evidence instead of trusted summaries.
+
+- Declared gate assertions must be satisfied before final review and completed
+  closure. New named evidence writes JUnit to the plan-bound
+  `.flow/results.xml`; execution cannot substitute another path. Reports are read once
+  from a stable, bounded workspace file without following symlinks.
+- Eval failures identify the evaluator, provider, host, or persistence boundary
+  that produced them. The release matrix uses repository-owned policy with enough
+  attempts to measure its 90% and 100% thresholds.
+- Exact-artifact canary results are derived from one OpenCode session lineage.
+  Release qualification seals all attempts, transcripts, provenance, canary,
+  artifact, and grader source into an immutable bundle that CI independently
+  reopens and regrades.
+- Publication requires the tag and packed artifact to match the release evidence.
+  npm and GitHub operations are bounded and idempotent, refuse conflicting bytes,
+  and recover through an exact draft without destructive asset replacement.
+- **Session v5 schema:** unchanged. Runtime commands and tools are unchanged.
+
+Install or update:
+
+```bash
+opencode plugin opencode-plugin-flow@8.1.3 --global --force
+```
+
 ## [8.1.2] - 2026-08-26
 
 Long Grok eval turns no longer fail at Bun's implicit five-minute fetch cutoff.

README.md

@@ -40,7 +40,7 @@ expensive, and it is overhead when it is not.
 Install the exact npm release through OpenCode:
 
 ```bash
-opencode plugin opencode-plugin-flow@8.1.2 --global --force
+opencode plugin opencode-plugin-flow@8.1.3 --global --force
 ```
 
 Omit `--global` for project scope. Version pins are exact and never update on
@@ -51,7 +51,7 @@ The equivalent manual project configuration is:
 ```json
 {
   "$schema": "https://opencode.ai/config.json",
-  "plugin": ["opencode-plugin-flow@8.1.2"]
+  "plugin": ["opencode-plugin-flow@8.1.3"]
 }
 ```
 

biome.json

@@ -8,7 +8,8 @@
 			"!bun.lock",
 			"!evals/results",
 			"!evals/canary",
-			"!evals/decisions"
+			"!evals/decisions",
+			"!evals/qualification/bundles"
 		]
 	},
 	"formatter": {