Agentic AI Atlasby a5c.ai
OverviewWikiGraphFor AgentsEdgesSearchWorkspace
/
GitHubDocsDiscord
iiRecord
Agentic AI Atlas · Release readiness — `staging` → `main`
page:docs-release-readiness-staginga5c.ai
Search record views/
Record · tabs

Available views

II.Record viewspp. 1 - 1
overviewarticlejsongraph
II.
Page JSON

page:docs-release-readiness-staging

Structured · live

Release readiness — `staging` → `main` json

Inspect the normalized record payload exactly as the atlas UI reads it.

File · wiki/docs/release-readiness-staging.mdCluster · wiki
Record JSON
{
  "id": "page:docs-release-readiness-staging",
  "_kind": "Page",
  "_file": "wiki/docs/release-readiness-staging.md",
  "_cluster": "wiki",
  "attributes": {
    "nodeKind": "Page",
    "sourcePath": "docs/release-readiness-staging.md",
    "sourceKind": "repo-docs",
    "title": "Release readiness — `staging` → `main`",
    "displayName": "Release readiness — `staging` → `main`",
    "slug": "docs/release-readiness-staging",
    "articlePath": "wiki/docs/release-readiness-staging.md",
    "article": "\n# Release readiness — `staging` → `main`\n\n_QA round run `01KXZE14FY9SSQQ8ENFVZVYXAD`, 2026-07-21. Analyzed staging HEAD `95a0b6bab`+ (branch moved during the round — a concurrent kip-sdk run was actively committing; see Caveats)._\n\n## Verdict: GO — `ci.yml` is fully green on the fixed staging sha\n\nThe bug fixes this release claims to deliver are **real and correct**, but the round found they shipped **structurally unguarded** — their tests either did not exist or did not run in CI, and one shipped fix (RC-1) was actively breaking every plugin-install CI lane. All of that is now closed on `staging` and **confirmed green by a real CI run against the release sha**.\n\n**CI confirmation (`ci.yml` run `29812297914` = SUCCESS, all jobs):** Docs QA ✓, Lint/Tests/Package ✓ (the newly-wired policy/tools/trust-core security suites), Observer Dashboard ✓, Workspace Coverage ✓ (Caveat 1 fixed). **Live Stack bp/predefined interactive claude-code/gpt-5.5 lane ✓** — the exact install-mode lane RC-1 was breaking now passes end-to-end.\n\nOnly by-design deferrals remain (owner's call), none release-blocking:\n- **kip-sdk CI (recommendation):** the largest part of the diff, but a **new/unreleased** package; its cross-OS workflow already exists untracked as the concurrent kip run's artifact — commit it. Not committed here to avoid clobbering that active run.\n- **RC-4 (recommendation):** `ci.yml` runs only on `pull_request`; add a `staging` push-trigger or branch protection so direct-push landings can't ship unguarded again.\n- **RC-3 (deferred, owner-approved):** external dispatcher ref hygiene, not product code.\n\n_Original Caveat 1 (transport-adapter Workspace Coverage TS2307) was **introduced by this release** (transport-adapter's proxy-attestation wiring added trust-core/policy-adapter deps; the isolated job built only atlas first) and masked by RC-4. Fixed in `26ff42325`: build the dependency closure in order. Reproduced locally (`rm dist` → identical TS2307) and confirmed green in the CI re-run above._\n\n---\n\n## 1. What this release claims to fix\n\n162 commits since the last main merge (`23a00eafd`, 2026-07-01). The volume is **inverted from the risk**: 192 of 408 changed files are the never-before-released `@a5c-ai/kip-sdk`, whose regression risk is near-zero regardless of coverage. The real release risk sits in ~10 small commits against already-shipped surfaces, clustered as:\n\n| Cluster | Risk | Claim |\n|---|---|---|\n| C1 policy authorization | blocker | Milestone A–E proof-based policy engine closed authorization-bypass defects (delegation-chain linkage, argv-evasion, attestation binding, quorum spoofing, fail-open signed-breakpoint, gate wiring). |\n| C2 policy activation | blocker | GATE-3 credential injection + authorization ALLOW path + code_executor invariant made activatable. |\n| C3 trust-core extraction | high | Trust primitives extracted to `@a5c-ai/trust-core`; extraction orphaned 4 imports and broke typecheck (shipped broken ~17 days, fixed today by `95a0b6bab`). |\n| C4 codex marketplace + hooks | blocker | `babysitter-codex` marketplace-addable (codex-format manifest); adapters-hooks stopped emitting the Codex-invalid `decision:\"noop\"` sentinel. |\n| C5 sdk bin + marketplace owner | blocker | Plugin hooks resolve `adapters-hooks` bin; root marketplace load fixed (owner object). |\n| C6 observe command | medium | `observe` command frontmatter quoting (unquoted `[...]` argument-hint broke load). |\n| C7/C8 kip-sdk | low | Internal debt closure + CLI live-demo defects in the new package. |\n\n## 2. Validation status per cluster (after adversarial refutation)\n\nVerdicts are **post-refutation** — every \"covered\"/\"partial\" claim was attacked, several downgraded. The recurring finding: **tests existed but did not run in CI.**\n\n| Cluster | Verdict (pre-fix) | Evidence |\n|---|---|---|\n| C1 | **uncovered** | Tests genuinely fail on revert (deleting the delegation-chain check → 4/11 fail; restoring old argv-matcher → 13/71 fail) — but `packages/adapters/policy` was in **no workflow** and absent from `adapters-build.cjs`; `ToolDispatcher.create` production wiring was deletable with 201/201 green. |\n| C2 | partial | Reverting `spawn-runner` gate3 wiring left **970/970** passing; the shipped test calls the resolver directly, never drives the runner. |\n| C3 | partial | Orphaned-import break was catchable by genty-core typecheck but survived 17 days because direct-push landings bypass PR `ci.yml` (RC-4); trust-core suite gated nowhere. |\n| C4 | **uncovered** | Zero tests. `scripts/sync-external-plugin-repos.mjs` (the marketplace generator users consume) had no test anywhere; the noop-sentinel guard was unexercised. |\n| C5 | partial | The load-bearing bin metadata + marketplace owner had no test; the bin fix (`bd11c5b1b`) also **caused RC-1**. |\n| C6 | **uncovered** | No test YAML-parses command frontmatter. |\n| C7/C8 | partial | Tests are solid and fail on revert, but kip-sdk runs in **no committed CI** (the only kip vitest workflow is untracked). |\n\n## 3. CI stack matrix — triage\n\n`ci.yml` (lint/build/typecheck/test) triggers only on `pull_request`; the July work landed via **direct push**, so **`ci.yml` last ran on staging 2026-06-30** and never exercised the release contents (**RC-4**). Root causes from the failing 2026-07-20 runs:\n\n| ID | Category | Blocker | Root cause |\n|---|---|---|---|\n| RC-1 | **product-regression** | yes | Every `bp/*` (install-mode) Live Stack lane fails at `npm install --global ./packages/adapters/hooks/cli`. `bd11c5b1b` gave `babysitter-sdk` an `adapters-hooks` bin that `@a5c-ai/hooks-adapter-cli` also declares → second global install EEXISTs. |\n| RC-4 | environment | yes | `ci.yml` never ran on the branch (pull_request-only + direct-push landings). |\n| RC-2 | flake | no | One wave had a truncated `build-all-dist` artifact (missing genty dist); self-corrected, but the workflow fail-opened, masking it as a product failure. |\n| RC-3 | environment | no | QA Dispatch checkout against non-resolving refs (external dispatcher). Deferred (owner-approved). |\n\n## 4. Gaps closed this round\n\nAll committed surgically (per-gap explicit `git add`, not a broad add — a concurrent kip-sdk run had uncommitted WIP in the shared tree) and pushed to `staging`.\n\n| Gap | Commit | Verification |\n|---|---|---|\n| RC-1 bin collision + install stderr | `92efd59bb`, `4d5a1451b` | `--force` de-collides the two providers; regression tests guard **both** the contract spec and the runner's executed install list — **fail on revert**. Documented user flow unaffected (installs only babysitter-sdk). |\n| Wire policy/tools/trust-core into CI | `7d00e7d3d` | Explicit build+typecheck+test steps in `ci.yml` + `adapters-build.cjs`. Confirmed green in CI: **`ci.yml` \"Lint, Tests, Package\" = success** (trust-core 46, policy 230, tools 206). |\n| Deletable gate wiring (C1/C2) | `83f3b24a8` | Real `ToolDispatcher.create` deny test (**fails if `create` neutered**) + spawn-runner gate3 credential-leak test (**fails if `gate3=undefined`**). Revert-verified. |\n| RC-2 fail-fast | `099a06b84` | Missing/truncated-dist signature now fails fast with a labelled message; legit flaky-postinstall tolerance unchanged. 4 tests. |\n| C4 marketplace generator + noop-sentinel | `44daa132c` | Codex-format marketplace test + noop-sentinel e2e — both fail on revert. |\n| C5 marketplace owner + sdk bin | `44daa132c` | Assert `.claude-plugin` owner + sdk bins — fail on revert. |\n| C6 command frontmatter | `44daa132c` | YAML-parses every `plugins/*/commands/*.md`; fails on unquoted argument-hint. |\n| G-WINPATH cross-platform paths | `ffe807c6d` | `getSessionFilePath` → `path.join`; 8 Windows-only test failures resolved (Linux CI was unaffected). |\n\n**Gates after remediation:** local `build:sdk && test:sdk && verify:metadata` → **2328 passed / 9 skipped**. Adversarial review of the 6 core commits → **89/100, zero gamed tests, no fallbacks, no weakened/skipped tests, lockfile untouched**; its one actionable finding (RC-1 test didn't guard the runner's executed list) closed by `4d5a1451b`.\n\n## 5. CI verification against the fixed sha (G-RC4-VERIFY)\n\nDispatched `ci.yml` and Live Stack via `workflow_dispatch --ref staging`:\n\n- **`ci.yml` `29809037087`** — \"Lint, Tests, Package\" = **success** (proves the newly-wired security suites + build + test:sdk pass on the release sha). Overall run = failure **only** from the pre-existing \"Workspace Coverage / transport-adapter\" job (Caveat 1) — no commit in this release touches transport-adapter.\n- **Live Stack vanilla lane `29809134002`** — claude-code/gpt-5.5 non-interactive **vanilla** lane = **success** (confirms the branch live-stack works and the RC-2 change is safe).\n- **Live Stack bp lane `29811363446`** — dispatched to exercise the RC-1 install path end-to-end (in flight at report time). RC-1 is otherwise proven at the unit level (revert-verified regression tests guarding both fix sites) and by static mechanism analysis.\n\n## 6. Tracked debts / deferred (owner-visible)\n\n- **Caveat 1 — `ci.yml` Workspace Coverage / transport-adapter QA surface fails.** Pre-existing, unrelated to this release; surfaced only because RC-4 kept `ci.yml` from ever running on the branch.\n- **RC-4 root fix (recommendation, not applied):** `ci.yml` runs only on PRs; direct-push landings never gate. Add a `push: [staging]` trigger or branch protection so this can't recur. Not applied here to avoid running the full matrix on every commit without owner sign-off.\n- **G-CI-KIP (recommendation):** kip-sdk runs in no *committed* CI. A complete cross-OS workflow (`.github/workflows/kip-conformance-cross-os.yml`) already exists **untracked** as the concurrent kip run's artifact — commit it. Not committed here to avoid clobbering another active run's WIP; kip is unreleased (low risk).\n- **G-C8-REG (recommendation):** kip `register --include/--exclude/--git-sha` flags have no direct test. Same kip territory; unreleased.\n- **RC-3 (deferred, owner-approved):** QA Dispatch non-resolving refs come from an external dispatcher, not product code.\n\n## 7. Caveats on the round itself\n\nA **concurrent orchestration run** (`kip-entity-linker`) was committing to `staging` throughout this round; HEAD moved several times and its uncommitted kip-sdk WIP shared the working tree. All audits recorded the sha analyzed; all commits were surgical to avoid bundling that WIP; audit subagents were read-only or restored their scratch edits. This is a standing hazard of overlapping runs on one branch, not specific to this release.\n",
    "documents": []
  },
  "outgoingEdges": [],
  "incomingEdges": [
    {
      "from": "page:docs",
      "to": "page:docs-release-readiness-staging",
      "kind": "contains_page"
    }
  ]
}

Shortcuts

Back to overview
Open graph tab