Document explicit tool generation and recovery

.github/workflows/pages.yml

@@ -15,7 +15,7 @@ concurrency:
   cancel-in-progress: true
 
 env:
-  AGENTCTL_COMMIT: "bb8fb0a5a28d876fee26c49c49a8a5bf1c8390e2"
+  AGENTCTL_COMMIT: "7dee64e1d1b6fd38c1883d0d235d712590c30fbe"
   NODE_VERSION: "24.4.1"
   PNPM_VERSION: "11.9.0"
   RUST_TOOLCHAIN: "1.88.0"

docs/execution/RELEASE_CONTAINER_REMEDIATION_20260906.md

@@ -256,3 +256,41 @@ The offered remediation ZIP is identical to the separately prepared clean export
 The actual uploaded artifact `10040098155` is 4,375,370 bytes with SHA-256 `c069d43f044e062cdf6f0e2eabcb171ecc29ce1e57fd3ad22009a5c94d93abf6`, matching Actions. Independent verification confirmed `.nojekyll`, all 94 HTML pages without retired branding or source boilerplate, exact clean source metadata and all 21 packages with every nested manifest/file digest. Every package matches its locally validated source package. The hosted remediation ZIP also matches the separate clean framework export byte for byte, including all 34 payload hashes.
 
 The artifact expires on 22 September 2026 at 03:47:25 UTC. Its actual ZIP, complete hosted log and Actions metadata remain under `hosted-bb8fb0a`; `verification.json` has SHA-256 `60ba75abe63693c493c5e25d35eb739f797bad2f2d6bdc0dc430505644cb24ba`. `site-clean-export-hosted-bb8fb0a.json` records the clean-export comparison. This subsequent ledger-only commit preserves tested documentation source `826071b`, framework pin `bb8fb0a` and all package bytes; only writing, Markdown and diff checks are repeated locally. Successful live preflight and full remediation remain separate gates; this source has not yet been deployed by this documentation update.
+
+## Approved revised-output deployment
+
+The user approved documentation PR8, merged as `f80553b2695b537b9e849e559c92b8c3370f1647`. Its tree matched reviewed evidence head `878560e`. [Normal Pages run 34187707168](https://github.com/opensourceops/opensourceops.github.io/actions/runs/34187707168) passed validation in 8 minutes 1 second and deployed in 12 seconds. The hosted browser suite passed 140 checks in 4.8 minutes with the unchanged generic mobile-search skip. The public source advanced from `0d4542c` to `bb8fb0a`.
+
+All 116 public files, including 92 HTML pages and all 21 downloads, matched the actual Pages artifact. All 346 nonhidden artifact files matched the reviewed PR build; the pinned official Pages uploader omitted only `.nojekyll`, which remained present in the validated PR artifact. Search and the mobile tutorial passed, and visible source boilerplate was absent. Report `deployed-bb8fb0a/deployment-evidence.json` has SHA-256 `8831ad090d0eb01e9d87dfd8b6bb623cb0a3fa154cf13a359da0dcc8216a930d`.
+
+The deployed tutorial exposed a second stale source sentence: it said both full-journey roles had 2,048 output tokens, although the analyzer used 2,048 and the implementer used 4,096. The preceding paragraph stated the correct distinction. This source-prose contradiction was recorded separately from the successful deployment and is corrected in the next source below. The earlier claim that the README fully agreed with the limits was therefore too broad; its 512-token preflight sentence had been corrected, but this second sentence remained.
+
+## Tool generation and runtime authority: framework 7dee64e
+
+Framework `7dee64e1d1b6fd38c1883d0d235d712590c30fbe` corrects the remaining role-limit sentence and adds the explicit boolean `providerOptions.toolStrict` for OpenAI and Azure OpenAI agents. Strict tool generation remains the default. The remediation preflight and implementer select `false` while retaining the complete exact path, content-pattern and length schemas. Runtime validation still rejects invalid arguments before their tool effect can dispatch, and final structured outputs remain strict. The selected option is persisted and changes affected recovery identities; recorded replay retains the recorded choice.
+
+The real CLI preflight passed at this source using `gpt-6-astra`, high reasoning and one exact pure echo with both multiline input objects. Its durable receipt records two requests, 1,208 total tokens, eight seconds and USD0.0176 estimated cost. The binary was a local native debug build, not a release package. An independent replay without the credential, after changing the instruction file and with macOS network access denied, produced zero fresh provider requests and effects and left the original state unchanged. Evidence is retained in `live-preflight-tool-strict/verification.json`. This establishes that bounded preflight and replay case; full two-role remediation, its resulting PR and release-image execution remain separate gates.
+
+The documentation branch starts from the approved merged Pages source and advances the framework pin, imported guidance and complete downloads together. Earlier source-specific passes and failures remain unchanged. This update prepares a draft PR and does not merge or deploy it. The local shared Cargo target was unavailable when work began, so source import, clean export and independent site checks proceed before the paired framework documentation gate.
+
+The shared-target framework generation and Clippy commands completed normally without interruption. Local verification then passed both components of the paired gate against the same clean source: `cargo xtask docs-verify` using the existing target and `pnpm verify` using pnpm 11.9.0 with `AGENTCTL_REPO` set to that checkout. The site verified generated freshness, writing, Mermaid, workflow pins, Markdown, spelling, Astro, all 94 HTML pages, routes, links, anchors and search. It checked 171 exact source objects and 17 remaining external URLs. Browser checks passed 140 cases in 2.6 minutes with the unchanged generic mobile-search skip. Logs are `framework-docs-7dee64e-attempt1.log` and `site-7dee64e-attempt1.log`; hosted verification will execute the unchanged combined `pnpm verify:agentctl` command.
+
+The complete remediation download is byte-identical to a separate clean export: ZIP SHA-256 `6ae2b1fa1a9d8737a075460bc30af21e2e9ec4eba49af1a09416b649c0f07827`, manifest SHA-256 `2df2f7b929ed03e797c28be9bf01258b3e6535194e11538466eadfbfdce12585`. All 34 payload hashes pass, the stale role-limit sentence is absent, and the preflight and implementer contain the explicit tool-generation setting. `site-clean-export-local-7dee64e.json` records the comparison. Hosted source and uploaded-artifact evidence follow the draft PR push; no full remediation or release publication is inferred from these checks.
+
+## Hosted tool-generation source proof
+
+[Hosted run 34195633694](https://github.com/opensourceops/opensourceops.github.io/actions/runs/34195633694) passed the unchanged combined `pnpm verify:agentctl` gate for documentation `2f2e7b03870899c9cc6f425e9d9eb8c005b15d99` and framework `7dee64e1d1b6fd38c1883d0d235d712590c30fbe`. Validation completed in 6 minutes 38 seconds. Browser checks passed 140 cases in 3.9 minutes with the unchanged generic mobile-search skip; all dedicated package checks passed. The hosted checkout verified 166 exact source objects and 22 remaining external URLs. Deployment and production Pages upload were skipped.
+
+The actual uploaded artifact `10043952664` is 4,377,533 bytes with SHA-256 `bba2fc3895a18e77ec92bdf1236499a403c19eeeb47c0d7008e2e21bd00b962a`, matching Actions. Independent inspection confirmed `.nojekyll`, all 94 HTML pages without source boilerplate or retired branding, exact clean source metadata and all 21 packages with every nested manifest and payload digest. Each package matches its locally validated source package. The hosted remediation ZIP is byte-identical to the separate clean framework export, including all 34 payload hashes, and the stale role-limit sentence is absent.
+
+The artifact expires on 22 September 2026 at 06:46:12 UTC. Its actual ZIP, complete hosted log and Actions metadata remain under `hosted-7dee64e`; `verification.json` has SHA-256 `d3b25f9d31ad4e0fb201a327b14bded8236f838dfea4e9e2563e75b3f3151514`. `site-clean-export-hosted-7dee64e.json` records the clean-export comparison. This subsequent evidence-only commit preserves tested documentation source `2f2e7b0`, framework pin `7dee64e` and all package bytes. Only writing, Markdown and diff checks are repeated locally for this ledger update.
+
+## Live candidate and publication reconciliation
+
+[Candidate remediation run 34195429813](https://github.com/Ompragash/agentctl-remediation-demo/actions/runs/34195429813) completed successfully using framework `7dee64e` and demo source `dfd8ba73af22eb9e121f0923e89d162c0b9bcf64`. It created [actual remediation draft PR10](https://github.com/Ompragash/agentctl-remediation-demo/pull/10), whose observed head was `458ea6480616ae8a5c9890646e466a577d762ca7`. The PR changes only `requirements.in` and `requirements.lock`; it remains an unmerged draft. Independent artifact, remote-tree, scan and durable-runtime reviews passed against that exact candidate source.
+
+The live durable receipt records three `gpt-6-astra` requests with high reasoning, 4,290 total tokens, two scoped file writes, 28 seconds of runtime usage and USD0.089960 estimated cost. Reasoning tokens are included once within output tokens; the estimate is not an invoice. The lease was reconciled. Remediation and eligibility both replayed without a credential and with container networking disabled, preserving recorded outputs and artifact digests with zero fresh provider requests or effects. SQLite integrity, persisted state, effect/result records, budget accounting and CAS hashes matched the inspections. Report `remediate-34195429813/live-verification.json` has SHA-256 `4dbfd5c0ef07ca230c12e3752d8e9e4fd7798ec78a09b0ac452f0c1815e3f556`.
+
+The trusted outer pipeline built and tested the original and changed application; each isolated application test run passed two tests. Rescanning against the same recorded vulnerability database removed the three targeted advisories, introduced no new findings and disclosed 270 unrelated residual findings. Those residual findings remain unresolved. Independent verification recomputed eligibility and matched the exact two-file remote PR tree. Report `remediate-34195429813/publication-verification.json` has SHA-256 `6ffd78a9f761f419a3640f41f63a46aa37d4f9bb3d4b7fead4bc13472b420d8c`.
+
+[Publication reconciliation run 34196095845](https://github.com/Ompragash/agentctl-remediation-demo/actions/runs/34196095845) passed with the paid jobs skipped before any steps. It revalidated the recorded publication artifact and reused the same branch, fingerprint, validated tree and draft PR with zero fresh provider requests. The verification report `reconcile-34196095845/verification.json` has SHA-256 `a35fd2e745557f72d3ee30f4a6703cd0c0145386cac64b85523e25859ac7cffd`. This verifies outer CI publication reconciliation. Runtime replay and selective recovery have separate execution evidence; no exactly-once GitHub delivery guarantee follows. Production publication and smoke testing of the published image remain separate operator steps.