feat(release): qualify narrow patches from preserved baseline evidence

.agents/plans/06-patch-release/README.md

@@ -0,0 +1,81 @@
+# Baseline-qualified patch releases
+
+This policy adds a limited path alongside full qualification. The operator
+explicitly authorized it on 2026-09-14. It does not change historical grades.
+
+## Eligibility
+
+A stable version must advance within a fully qualified baseline's major/minor.
+Every candidate is compared directly with that baseline's measured source, never
+just the previous patch. The baseline tag must resolve to the recorded commit,
+contain the measured source commit, and be an ancestor of the candidate.
+
+Runtime source, lockfiles, dependencies, package exports, package scripts,
+engines, build configuration, and all guides other than flow-run must remain
+unchanged. Package metadata may change only its version. Only enumerated release
+scripts/workflows, offline evals, tests, plan records, documentation, and .gitignore
+may otherwise differ. Unknown paths fail. Only adding the sealed patch-baseline directory to formatter exclusions is
+allowed in biome.json; all other settings must match. The build and packaging helpers are
+outside the allowlist. A new category requires an explicit reviewed policy change.
+
+The only shipped guide exception is a reviewed flow-run clarification. Its exact
+content digest and rationale must be recorded. The reviewer must confirm that it
+clarifies an existing obligation without changing permissions, workflow gates,
+validation, review requirements, or closure guarantees. Code cannot infer that
+semantic judgment from a path or a patch version. A content change invalidates
+that approval. This is a maintainer review control, not a sandbox against someone
+who can rewrite the policy and its approvals together.
+
+## Evidence
+
+A record at evals/qualification/patches/VERSION.json uses schemaVersion 1 and kind
+baseline-qualified-patch. It contains the candidate's exact tarball and unpacked
+manifest hashes, its version, the baseline artifact identity, tag commit, canary
+path and regraded bundle hash, approvedBy, rationale, and any guidance approval.
+The PatchReleaseSchema in scripts/patch-release.ts is the executable format.
+
+The baseline must independently pass the full verifier at release time. Existing
+seven-day campaign and 72-hour canary limits remain unchanged. A patch record
+cannot stand in for a fully qualified baseline, extend freshness, or invent runs.
+An expired baseline blocks; it never starts a paid campaign automatically.
+
+The candidate has no new model eval or live model canary. Required assurance is
+clean source, strict diff eligibility, matching package bytes, review of the
+clarification, deterministic checks, replay, package smoke, provider-free pinned
+OpenCode smoke, and green CI. These checks do not establish model behavior on the
+new candidate. This reduced assurance must be visible in release notes.
+
+Offline verifier changes may require resealing the original retained baseline
+campaign under current verifier authority. Keep original provider evidence and
+seals unchanged; place the new seal alongside them in a separate bundle directory.
+Only that new directory is passed to patch verification, avoiding ambiguous seals.
+
+## Publication
+
+The release workflow chooses --patch only when the version's explicit record
+exists; otherwise it uses --canary and the original full route. A malformed patch
+record fails; it never silently falls back. Both flags together are rejected.
+
+release.ts init and resume independently recheck the same evidence choice. The
+persistent release record binds the patch file's digest through bundleSha256 and
+stores its path in the optional patch field. Existing full records stay valid.
+Candidate notes identify the policy, prior version, prior package/bundle/canary,
+and state that prior passes are not candidate measurements. Recovery still checks
+exact payloads, Git refs, remote assets, and npm integrity before mutations.
+
+No --skip-evals flag, environment bypass, altered old report, or fake candidate
+canary is used. Updating this policy does not itself publish or authorize spend.
+
+## Current candidate
+
+8.3.1 contains the reviewed baseline capture/inventory clarification from PR #75.
+The independent reviewer already required that inventory. Core source, dependencies,
+build settings, and review authority are unchanged. PRs #73/#74 improve release
+recovery and spending guards; #76 adds offline reporting. The intended release
+notes must describe the spending guard as a top-level dispatch limit, not a token
+or billing ceiling. The 8.3.0 matrix's 76/76 result belongs to 8.3.0 only.
+
+The regression suite checks version boundaries, protected paths, dependency
+changes, guide hash substitution, dirty source, changed runtime, moved baseline
+tags, stale evidence, and mismatched seals. Existing publication recovery tests
+continue to exercise unchanged transport convergence. No paid runs are authorized.

.agents/plans/06-patch-release/verification.md

@@ -0,0 +1,30 @@
+# 8.3.1 patch verification
+
+The original 8.3.0 campaign was independently regraded offline into the new seal
+qb1-c61abd32788d84b1950e7908cf3b96bc57e4dad1a6af26c950a6d3124bd2e6a4.
+Original reports, model transcripts, canary, and seals are unchanged. The copied
+content-addressed baseline bundle retains its own verifier source evidence.
+
+Candidate tarball SHA-256:
+29c7a3c1a4a652b5e3b95f9966d502556d944317bc73b820e2e048067e3df14c.
+Candidate unpacked manifest SHA-256:
+53f731df696d1a20c5ed181580ef841d73eecffab32820725291242a95e200d2.
+
+The real release initializer verified eligibility, package identity, guidance
+content, and full baseline evidence, then generated the durable local release
+record and notes. This was initialization only; no publication resume was run.
+The draft output lives under ignored .release-artifacts/8.3.1-readiness/.
+
+Focused release tests passed (10), existing metadata/documentation tests passed,
+all 13 gated cassettes reproduced, package smoke passed, actionlint passed, and
+pinned OpenCode 1.18.6 provider-free smoke passed (21 tests). These do not run
+models. The canary readiness dry-run correctly says the candidate has no canary;
+the explicit patch policy accounts for that absence.
+
+An initial initializer invocation used an abbreviated commit and was rejected by
+the release-record schema. Repeating it with the full commit passed. No evidence
+was weakened or replaced to resolve this input error.
+
+The patch baseline retains the existing 72-hour canary and seven-day campaign
+freshness limits. Before a later publication attempt, verify it has not expired.
+CI and PR review remain required before merging/releasing the new policy.

.github/workflows/ci.yml

@@ -113,3 +113,43 @@ jobs:
 
       - name: Run live OpenCode smoke
         run: bun run smoke:live
+
+  patch-qualification:
+    runs-on: ubuntu-latest
+    timeout-minutes: 10
+    steps:
+      - name: Check out full release history
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+        with:
+          fetch-depth: 0
+      - name: Set up Node
+        uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7.0.0
+        with:
+          node-version: 24
+      - name: Set up Bun
+        uses: oven-sh/setup-bun@0c5077e51419868618aeaa5fe8019c62421857d6 # v2.2.0
+        with:
+          bun-version: 1.3.14
+      - name: Verify patch candidate offline when declared
+        env:
+          BASE_SHA: ${{ github.event.pull_request.base.sha || github.event.before }}
+        shell: bash
+        run: |
+          set -euo pipefail
+          version="$(node -p "require('./package.json').version")"
+          patch="evals/qualification/patches/${version}.json"
+          if [ ! -f "$patch" ]; then
+            echo "No baseline-qualified patch declared. Full qualification remains required for release."
+            exit 0
+          fi
+          if [ -n "$BASE_SHA" ] && git diff --quiet "$BASE_SHA" HEAD -- "$patch" package.json; then
+            echo "No new patch release declaration; candidate qualification deferred until release preparation."
+            exit 0
+          fi
+          bun install --frozen-lockfile
+          bun run build
+          mkdir -p .release-artifacts/patch-ci
+          bun pm pack --destination .release-artifacts/patch-ci
+          bun scripts/release.ts init .release-artifacts/patch-ci/record \
+            --artifact ".release-artifacts/patch-ci/opencode-plugin-flow-${version}.tgz" \
+            --patch "$patch" --commit "$GITHUB_SHA"

.github/workflows/release.yml

@@ -157,16 +157,20 @@ jobs:
             exit 1
           fi
 
-      - name: Verify independently regraded qualification bundle and fresh canary
+      - name: Verify full or baseline-qualified patch evidence
         timeout-minutes: 5
         shell: bash
         run: |
           set -euo pipefail
           version="$(node -p "require('./package.json').version")"
           tarball="$(ls opencode-plugin-flow-*.tgz)"
+          evidence=(--canary "evals/canary/${version}.json")
+          if [ -f "evals/qualification/patches/${version}.json" ]; then
+            evidence=(--patch "evals/qualification/patches/${version}.json")
+          fi
           bun run scripts/release.ts init .release-state \
             --artifact "$tarball" \
-            --canary "evals/canary/${version}.json" \
+            "${evidence[@]}" \
             --commit "${GITHUB_SHA}"
 
       - name: Persist release record before publication

CHANGELOG.md

@@ -2,6 +2,27 @@
 
 One short entry per release, written for users deciding whether to upgrade.
 
+## [8.3.1] - 2026-09-14
+
+Clarify review handoffs and make release operations recoverable.
+
+- Capture the feature baseline before edits and include its inventory in review.
+  Recover retained observations on resume and identify missing evidence explicitly.
+- Resume publication from a durable release record without repeating uncertain
+  GitHub draft creation. Verify existing assets and npm integrity on recovery.
+- Require persistent authorization for supported paid eval and canary launchers.
+  Consumed top-level dispatches survive failure and restart; this is not a token
+  or billing ceiling.
+- Report review observations, missing evidence, and attributed finding categories
+  offline. Advisory findings may omit optional evidence without losing the review.
+- Add a restricted baseline-qualified patch route. This candidate has no new live
+  model eval or canary; prior 8.3.0 results describe 8.3.0 only. Offline checks and
+  reviewed scope support this release, without claims of improved model accuracy
+  or measured cost savings.
+- Schema impact: Session v5 schema, runtime tool schemas, and package exports are unchanged.
+
+Upgrade with `opencode plugin opencode-plugin-flow@8.3.1 --global --force`.
+
 ## [8.3.0] - 2026-09-13
 
 Flow preserves native model settings across continuation, keeps full sessions

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@8.3.0 --global --force
+opencode plugin opencode-plugin-flow@8.3.1 --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@8.3.0"]
+  "plugin": ["opencode-plugin-flow@8.3.1"]
 }
 ```
 
@@ -62,7 +62,7 @@ For an explicit reviewer model, use OpenCode's plugin tuple options:
   "$schema": "https://opencode.ai/config.json",
   "plugin": [
     [
-      "opencode-plugin-flow@8.3.0",
+      "opencode-plugin-flow@8.3.1",
       { "reviewer": { "model": "provider/model", "steps": 80 } }
     ]
   ]

biome.json

@@ -9,7 +9,8 @@
 			"!evals/results",
 			"!evals/canary",
 			"!evals/decisions",
-			"!evals/qualification/bundles"
+			"!evals/qualification/bundles",
+			"!evals/qualification/patch-baselines"
 		]
 	},
 	"formatter": {

docs/release-qualification.md

@@ -29,16 +29,15 @@ This page owns release thresholds, candidate freezing, and publication order.
 | `project-gate-discovery` | ungated | Report-only until two provider baselines show whether planning selects the explicit whole-repository command over a narrower script. |
 | `task-risk-lenses` | ungated | Measures whether the manager supplies the relevant full review questions. It does not measure defect detection or false blockers. |
 
-Offline verification fixes do not require new provider runs when the measured
-package bytes and frozen case policy remain unchanged. Qualification preserves
-the original report and evaluator identity, verifies that identity against the
-exact recorded Git commit without executing historical code (that commit must
-be available in local Git history), and retains those
-execution sources separately from the current verifier sources. The current
-verifier independently recomputes every retained outcome before publication.
-Missing or mismatched execution sources, changed outcomes, and altered package
-bytes remain failures. Canary review retries may use multiple reviewer sessions
-when all are linked to the same observed manager with consistent identities.
+Offline verifier fixes may reuse runs only for unchanged package bytes and frozen
+case policy. Current code regrades retained outcomes without executing historical
+code; execution sources must match the recorded Git commit. Missing sources or
+changed outcomes fail. Canary retries must retain consistent manager/reviewer identity.
+
+Eligible patches may instead use the [baseline-qualified patch policy](../.agents/plans/06-patch-release/README.md).
+This explicitly permits no new model eval or canary, binds the new artifact, and
+labels prior results as baseline evidence only. Full qualification remains the
+default; patch version numbers alone grant no exception.
 
 A new scenario needs an explicit release-policy decision. Any required canonical
 case missing from the report fails qualification.
@@ -72,7 +71,8 @@ Finish or close active sessions before changing Flow versions in either directio
 - **Freeze on the public surface** while the guarantees are being measured: tools,
   commands, guides, agents, and the Session v5 shape. Additive optional fields are
   allowed; removals and renames are not.
-- **No release** without a sealed V2 qualification bundle and fresh canary.
+- **Full qualification** requires a sealed V2 bundle and fresh canary; eligible
+  baseline-qualified patches use the explicit policy above.
   The bundle retains every attempt, transcript, grader source, and exact artifact
   needed to reproduce its decision. Release metadata independently regrades those
   bytes before publication and derives the provider-count evidence table added to