Prime Agent v0.8: five-stack integration tracker

[sethkarten]

Scope

This issue tracks the v0.8 integration across Core, MCP, Release, Prompts, and ACP/Evaluation. Upstream prerequisites remain separate: #838, #850, #851, and #852.

Integration gates

• Upstream prerequisites feat(coding-agent): edit queued messages in place and preserve the queue on interrupt #838 / fix(coding-agent): report accurate worker lifecycle and hide stopping workers #850 / fix(coding-agent): finalize timed-out worker stops instead of stranding registrations #851 / fix(coding-agent): self-heal stale worker registrations on resume #852 are merged and verified.
• Core has a remotely available, reviewed common integration ref.
• MCP, Release, Prompts, and ACP changes are consolidated into reviewable v0.8 PRs.
• Cross-stack integration and required checks are green.
• Approved evaluations pass with immutable evidence.
• Both repositories and final release assets are verified.
• Human-ready new v0.8 PRs are available for review.

Current human-ready new v0.8 PRs: none.

Current stack status

Stack	Tracker / status	References
Core	Reviewed remote component refs are available. The common Core ref is local and blocked on 11 monitor failures; it is not remote or reviewed.	Remote reviewed: C01 8159dbbb157af172f5d15501acc424d0b3b267df; C02 7837d6bc59fe2cb21330501569bba88174693844; C03 295c7aaf1b2418abe1d0030ed2637c9cb264cd9b; C04 9d4bb90c0158108b483aa5e0cdf9b32559263407; C05 trust 932f5f55eb54e30d32afab5b63b2078d2e385871; C06 6bf86d23bb92b9689cbf1b050cdc00da53c95e08; C07 6e2f93162bdf273ea02b8ff38c070811ce10b668; C08 9891e5d98f53d17cdbff9337af9386fd4d64b00f. Local blocked common ref: 894d1b5f82e511a2f6f5e8736b3c3f486a1f85ce.
MCP	Draft tracker #1175. Remote backup / partial refs are retained for recovery, not presented as a human-ready v0.8 PR.	Backup S01 39c539085499e5dc38bf64b08b4ba247a6d5c9e2; partial M01 ab39d1aa8719ab07e767cf584f9cf03b48aedb2a.
Release	Draft tracker #1176. Listed refs are remote references; no new human-ready v0.8 PR yet.	d0187164768c2ea54e48de7ca8573b182e0e31e8; bebbe039e3e4956bf2a1868f54559f143e8fa400; a7def5acfb258102be4a445714d0199d683c80e3.
Prompts	Draft tracker #1177; its head is green. Core dependency replay remains.	Head 64b72cce719c0b57cd67be1f029a8a22d3c837bd.
ACP/Evaluation	Draft tracker Prime #2310 plus existing Prime #805/#806 and Verifiers #2285/#2286. Reviewed remote backup refs are available; artifact and evaluation gates remain.	d8db5fb81f1a1fa0509c320ab1da01950ad83ce1; 4673116c6debf5fe75650784cdefe9178078fc5b; 170b95c5a2a85a571240fd3fb5b22ad792997812; 1073f7f3a21c9c01b7ec67dcc2669287ef6c0097; 490da85bee521f1805cc4a7ccacad14d592a43d5.

Next updates

• Keep prerequisite work in feat(coding-agent): edit queued messages in place and preserve the queue on interrupt #838, fix(coding-agent): report accurate worker lifecycle and hide stopping workers #850, fix(coding-agent): finalize timed-out worker stops instead of stranding registrations #851, and fix(coding-agent): self-heal stale worker registrations on resume #852.
• Resolve the Core common-ref monitor failures before treating it as an integration candidate.
• Promote only validated consolidated work to new v0.8 PRs, then update the gates above.
• Record approved evaluation and release evidence before marking v0.8 complete.

[sethkarten]

Prompts checkpoint — reviewed draft ref green

• Remote reviewed code ref: v080/prompts at 64b72cce719c0b57cd67be1f029a8a22d3c837bd (draft PR [v0.8 Prompts P01] preserve typed system-prompt provenance #1177, still draft).
• Reviewed stack: accepted C07 prerequisite; P01 typed built-in/custom/unknown prompt provenance; P02 built-in-only clear-direct-prose block replayed on the reviewed P01 chain.
• Validation: latest [v0.8 Prompts P01] preserve typed system-prompt provenance #1177 CI is green (15 successful checks, 1 skipped); local replay and successor diffs were clean. Initial isolated-worktree Vitest was unavailable before remote CI because dependencies were absent; remote CI subsequently passed.
• Blocked / not ready: [v0.8 Prompts P01] preserve typed system-prompt provenance #1177 is not Alex/Sebastian-ready. It must remain a draft until the complete v0.8 Core stack and post-upstream-replay integration are coherent. No backup ref or prerequisite is represented as human-ready.

[sethkarten]

ACP / Evaluation checkpoint — reviewed candidates backed up; release/eval blocked

Remote reviewed candidate refs (all exact non-force backups, not human-ready):

• Prime feat(coding-agent): keep ACP daemon sessions resident #805: v080/acp-805-reviewed → d8db5fb81f1a1fa0509c320ab1da01950ad83ce1
• Prime feat(coding-agent): expose ACP quiescence metadata #806: v080/acp-806-reviewed → 4673116c6debf5fe75650784cdefe9178078fc5b
• Verifiers #2285: v080/acp-2285-reviewed → 170b95c5a2a85a571240fd3fb5b22ad792997812
• Verifiers #2286: v080/acp-2286-reviewed → 1073f7f3a21c9c01b7ec67dcc2669287ef6c0097
• Negative conformance evidence: v080/acp-conformance-reviewed → 490da85bee521f1805cc4a7ccacad14d592a43d5

Local deterministic evidence: all five frozen worktrees are clean and independently reviewed. Focused suites passed P1 15/15, P2 19/19, V2 22/22, V3 54/54, V4 12/12. P2/V3 canonical fixture bytes match SHA-256 cacde7827aadf186db2ce1af1ea6f3b6d109504fa94945633bd9c0d96106b882.

Blocked WIP / not human-ready: these are divergent reviewed source candidates pending an authorized fresh merged-base replay. The Verifiers harness remains pinned to an approved 0.7.0 artifact while P2 source is 0.7.1; there is no approved immutable 0.7.1 artifact version/URL/SHA/provenance tuple. No live/paid evaluation is authorized.

Before readiness: verify post-merge successor ancestry and replay, bind the exact immutable producer artifact, then run the separately approved GSM8K + long-horizon plan with a named model/revision, numeric cost cap, frozen corpus/seeds, and retained raw traces/usage/billing evidence. end_turn remains transport-only, never correlation evidence.

[sethkarten]

MCP checkpoint — reviewed partial refs; common Core WIP blocked

Issue context: #1182

Remote reviewed refs (partial only): v080/mcp-s01-local = 39c539... (reviewed partial MCP-S01); v080/mcp-m01-contract-local = ab39... (reviewed partial M01). Neither is human-ready or final M01.

Tests: S01 runner unavailable. M01 probe passed 6/6; the other dependency test was unavailable.

Blocked WIP: final M01 awaits Core C05 production trust capability and a common reviewed C05-trust + C06 + R01-H lineage. M02/M03 do not start. Current common Core WIP is local and blocked by 11 monitor failures. No backup ref is human-ready.

[sethkarten]

Release checkpoint — REMOTE REVIEWED CANDIDATES; blocked/WIP

REMOTE REVIEWED CANDIDATES (non-force pushed; not human-ready/prerequisites):

• v080/release-rel01-immutable-manifest-candidate = d0187164768c2ea54e48de7ca8573b182e0e31e8
• v080/release-rel03-immutable-publication-candidate = bebbe039e3e4956bf2a1868f54559f143e8fa400
• v080/release-s01-freeze-evidence-candidate = a7def5acfb258102be4a445714d0199d683c80e3

Validation: exact git ls-remote --refs SHA matches; source worktrees clean. Prior focused tests: REL01 24/24, REL03 26/26, S01 diff checks.

BLOCKED/WIP: refs are divergent; no coherent replay or PR update (#1176 untouched). Final paired Prime Agent + Verifiers SHA is absent, as are clean-room/rebuild/fresh-install/eval/raw evidence and S01 approval.

Readiness: not ready. Preserve these refs; replay parent-first only after final merge ancestry.

[sethkarten]

ACP draft CI checkpoint — Verifiers successor reviewed; Prime remediation in progress

Verifiers Draft #2312 remains Draft / not human-ready. Its original reviewed head
1073f7f… had a causal Ruff F811 duplicate GSM8K test. The isolated reviewed
successor is now backed up at v080/acp-2286-ci-fix:
1cd51755702720e2f08667815c211b741734fc9c.

Independent review accepted that successor as a narrow CI fix only: it removes
exactly the second byte-identical 27-line test_prime_agent_solves_gsm8k block,
leaving the canonical GSM8K coverage and all ACP correlation code/fixtures
untouched. Local non-live checks reported Ruff pass and 68 focused tests passing.
The draft head is intentionally unchanged pending the tracked P1/P2 replay topology.

Prime Draft #1183 remains Draft / not human-ready. CI failure diagnosis shows
P1 resident-daemon lifecycle regressions (worker adoption/restart cleanup) rather
than P2 producer-correlation failure. A P1-only successor is under review/remediation;
P2 remains frozen.

No live/paid E2E was triggered or controlled. Artifact provenance and final
merged-base replay remain blockers; end_turn remains transport-only.

[sethkarten]

Safety correction: automatic live E2E was canceled; no live result is claimed

Draft #2312's internal branch automatically triggered V1 live E2E with secrets.
That is outside the approved evaluation gate. Parent canceled GitHub Actions run
31449081434; its live job 93649480332 is now cancelled. It started
Run live v1 E2Es at 01:22:25Z and was canceled at 01:31:11Z.

This cancellation is recorded only as a safety event, not as evaluation
output. Do not infer a pass/fail, score, cost, or artifact validity from it.
The draft head is frozen at the original reviewed SHA; the V3 Ruff-only fix remains
backup-ref-only. No draft head update or CI retrigger is allowed until a reviewed
workflow-level safe skip mechanism is authorized, or an explicit live-evaluation
approval supplies model, budget, artifact, corpus, and retention controls.

[sethkarten]

Checkpoint: remote reviewed ref v080/core-common-c05-c06-r01-reviewed now points to 35c71c45e6f5e21d91ad2793f24b30f329f3a897 (35c), with draft PR #1186. This is the reviewed coherent MCP-prerequisite substack only, not complete Core and not human-ready. Remaining local/upstream work: C01–C08 integration blockers plus required upstream and post-merge replay; reviewed evidence remains monitor 80/80 and matrix 141/1.

[sethkarten]

ACP draft CI follow-up — P1 remediation timebox reset; no partial fix published

Draft #1183 remains Draft / blocked / not human-ready. The CI failures are
causal P1 resident-daemon lifecycle failures, not P2 producer correlation.

A scoped P1 remediation investigated the custom-socket update-restart regression.
It found an inherited daemon-worker environment role could misclassify a spawned
supervisor, then verified an env-scrub helper unit path. However, the bounded
4606 regression still failed: after the replacement session worker listened, it
was terminated around the recovery deadline and recovery retried, so the
supervisor never reached the readiness/custom-socket handshake. A 60-second
coordinator startup allowance did not fix this deeper readiness-cycle cause.

To avoid publishing a deceptive partial change, the P1 worktree was reset exactly
to reviewed d8db5fb81f1a1fa0509c320ab1da01950ad83ce1; no P1 CI-fix commit or
backup ref was published. The complete discarded experimental patch is retained
outside repositories as session evidence (SHA-256
8f7d04c2db4448390c90cee60913f6d1e722df980781159f1fd50faae164e228).

A fresh deeper lifecycle owner/review plan is required before #1183 can advance.
P2 remains frozen. No live/paid evaluation or PR-ready claim is made.

[sethkarten]

P1 recovery forensic follow-up — do not cherry-pick later passivation fix onto #1183

Read-only local commit forensics evaluated later commits 59ef81556 /
ff0742ae1 (“defer processless owned worker recovery”) as a possible remedy for
the P1 recovery cycle. They are not a minimal compatible fix for frozen
#1183/P1 d8db5fb.

• The two later patches are effectively identical for daemon/passivation content,
  but git apply --check on frozen P1 fails in daemon-supervisor.ts; their new
  restart-passivation test does not exist at frozen P1.
• The later change assumes a prior processless/passivated worker-descriptor
  protocol: PID-less owner roots become dormant and only owner attach+env wakes
  them. Frozen P1 requires descriptor PID, rejects PID-less descriptors, forces
  valid records into recovery at startup, and marks dead no-env owned workers
  failed.
• The full prerequisite series also changes stop/adoption/wake/archive semantics,
  so literal cherry-pick/reimplementation is high-conflict and unsafe.

Decision: no partial P1 fix was published. #1183 remains Draft/blocked. A
future fix must either integrate/rebase the complete passivation+cleanup series
with end-to-end 4603/4606/4600 and owner-reattach validation, or implement a
d8-native correction around its persisted launch-env model. No live/paid action
or external mutation was performed for this analysis.

[sethkarten]

Material checkpoint: #838 is now landed on main at 71ca6cfd (squash-equivalent tree/diff to eb16eacc). #850 must restack/fix on post-#838 main; #851 is currently not a descendant of #850, and #852 is a descendant of #851, so the lifecycle stack still blocks a full Core replay. Core common PR #1186 remains DRAFT and is not human-ready.

[sethkarten]

v0.8 integration checkpoint — upstream lifecycle and Core/MCP holds

• Upstream fix(coding-agent): report accurate worker lifecycle and hide stopping workers #850 was updated by its contributor to 749831dd9dd1a0b35708513390080030878d9e4f on current main. A separately reviewed non-force prerequisite backup exists at v080/prereq-850-worker-lifecycle-restack-reviewed / 0026c8adc13d29fd9efcb52b28fe2adbb9dbb7a0, but it is now HOLD, not a ready replacement.
• New fix(coding-agent): report accurate worker lifecycle and hide stopping workers #850 CI run 31454512036 exposes four real lifecycle regressions: two ENG-4602 catch-up cases and two ENG-4677 retained/cached snapshot cases. The correction must distinguish supervisor-owned cached completion from actual unavailable-worker routing and must preserve the schema digest coverage. fix(coding-agent): finalize timed-out worker stops instead of stranding registrations #851 work remains local and cannot advance until the corrected fix(coding-agent): report accurate worker lifecycle and hide stopping workers #850 parent is accepted; fix(coding-agent): self-heal stale worker registrations on resume #852 has not been replayed.
• Draft Core [v0.8 Core WIP] bind project trust, recovery, and typed host requests #1186 remains unchanged at 35c71c45e6f5e21d91ad2793f24b30f329f3a897. The later local cfff84ed0e3005f4a4dd309f3ccdd3f308d94993 is BLOCKED: rejected optional-context commit 268bdf9... is in its ancestry, runtime unary handler calls still need fail-closed context guards, and the final line must be rebuilt cleanly after fix(coding-agent): report accurate worker lifecycle and hide stopping workers #850→fix(coding-agent): finalize timed-out worker stops instead of stranding registrations #851→fix(coding-agent): self-heal stale worker registrations on resume #852.
• MCP M01/S01/M02 remain accepted locally (8aafc1aa..., 909efe7a..., 0ce8570491f6188e11f6ee1d476dd3ff353f6d14). M03 is paused with no accepted candidate. Its required Core runtime-declaration snapshot is being refactored to reuse the single M01 parser and opaque C05 admission; the first f9f487... attempt is rejected for a forgeable admission wrapper, nondeterministic ordering, and parser hardening gaps.

No contributor branch was mutated, no force-push occurred, and no live/paid evaluation or release action ran. Ready for Alex/Sebastian: none.

[sethkarten]

v0.8 integration checkpoint — worker lifecycle prerequisite #850 validated

The corrected #850 prerequisite is now accepted at a6db0246b52cea74f2f675fea95ac43c9842d39a.

• GitHub CI run 31456400058: build/check, all three coding-agent shards, process smoke, kernel, package suites, and aggregate all passed.
• CodeQL run 31456396972: all analyses passed.
• Reviewed backup v080/prereq-850-worker-lifecycle-restack-reviewed was fast-forwarded non-force and remote-verified at the exact accepted SHA.
• Draft #1195 is only the clean CI/collaboration surface for this upstream prerequisite. It is not a v0.8 deliverable and is not being presented for human merge review.
• Retained CI log SHA-256: e7e612f97f200a5faed6ae7bfbd4f789827434bdf4869493200f54ef1d402e68; run metadata SHA-256: 88651a929b6dc03953fae50de2ada3e695ab82a959c24040d8c3ad8242deba0f.

Next: rebuild #851 on this exact base and close the missing-processStartId PID-reuse/SIGKILL safety gap before #852. In parallel, the MCP runtime declaration snapshot remains isolated and unpublished while one test-type import is corrected; M03 remains paused. No v0.8 stack is human-ready yet, and no live/paid evaluation or release action has been taken.

[sethkarten]

v0.8 prerequisite checkpoint — #851 remote validation found one lifecycle P0

• Draft validation surface: [WIP prerequisite] finalize worker shutdown with PID identity fencing #1202 at exact head f7ccf5e74d4d59d007ac889b3385be8eed87d798.
• Clean GitHub process smoke passed, as did CodeQL, kernel, agent-core/AI/TUI, and coding-agent shards 1/3 and 3/3. This supersedes the locally contaminated/overlapping process attempts; those remain invalid evidence.
• Build/check exposed three narrow inferred test-fixture type errors.
• Shard 2 also exposed a stale ENG-4602 harness registration and a real stop-finalization retry defect: a cleanup retry incremented stopRevision, fenced its own next attempt, and could strand a tombstoned worker registration.
• A sole writer is applying the narrow correction: retain the exact tombstone generation for finalizer retries (recoveryCleanup semantics), preserve the relaunch assertion, strengthen the transient cleanup regression, register the 4602 fixture correctly, and repair only the three fixture typings.

#1202 remains Draft/prerequisite-only and is not human-ready. #852 has not started. MCP M03 also remains held under final protocol/security remediation; no v0.8 stack is ready for Alex/Sebastian yet.