Track Pylon fork parity and upstream overlap

[rynfar]

This is the fork-side tracker for Pylon-specific Prime Agent integration work. The Pylon application-side umbrella is pylon-code/pylon#114.

Foundation

• Keep main as an exact daily mirror of PrimeIntellect-ai/prime-agent:main — established by PR chore(pylon): establish fork governance and sync #2 and verified at d60fab8a76d9c169f945341f0ee3bde21903bb55.
• Merge upstream into pylon through reviewed pull requests — first landed in PR chore(pylon): merge Prime upstream through d60fab8a #5 through d60fab8a76d9c169f945341f0ee3bde21903bb55.
• Record every overlap decision in .pylon/features.yaml and .pylon/upstream-review.md — established and exercised by PR chore(pylon): merge Prime upstream through d60fab8a #5.
• Configure a narrowly scoped Pylon GitHub App token for draft PR creation; keep mirror pushes on GITHUB_TOKEN.
• Define deterministic fork artifacts in Build deterministic Pylon Prime release artifacts #28, then publish attested preview/stable channels in Publish attested preview and stable Pylon Prime releases #29; do not restore inherited upstream release workflows.

Pylon reliability and delivery sequence

Pylon consumes the installed package's public SDK and detached-daemon API as its primary connector. ACP remains a tested degraded fallback only where native daemon mode is unavailable or unsafe. The Pylon application checklist is pylon-code/pylon#114.

• Ship correlated prompt ownership, immutable snapshot transfer, authoritative owned-session cleanup, bounded public ingress, and generation-scoped post-attach capability proof through PRs fix(coding-agent): preserve correlated prompt ownership #12, fix(coding-agent): keep snapshot and recovery failures isolated #14, fix(coding-agent): bound public daemon client JSONL ingress #16, and feat(sdk): prove negotiated daemon session capabilities #18.
• Finish mixed-version snapshot catch-up isolation in fix(coding-agent): finish mixed-version snapshot catch-up isolation #21 / PR fix(coding-agent): isolate snapshot catch-up failures #19, merged as 70b65305c, with exact-head maintainer approval and corrected ledger attribution.
• Land deterministic fork artifacts (Build deterministic Pylon Prime release artifacts #28 / PR feat(release): build deterministic Pylon artifacts #32 at fully green exact head 7d4f4a753, awaiting approval) and protected preview/stable publication with keyless provenance (Publish attested preview and stable Pylon Prime releases #29).
• Give Pylon signed distribution detection (pylon#193) and opt-in side-by-side install/update/switch-back (pylon#194) while preserving stock Prime.
• Add capability-gated recoverable owned-session adoption in Add capability-gated adoption for disconnected client-owned sessions #27 for Pylon #84, without prompt replay or duplicate settlement.
• Add caller-owned recovery environment and observable owned cleanup in Isolate caller-owned session recovery environment and cleanup #33 for Pylon #199. Final native multi-instance enablement also waits for Add capability-gated adoption for disconnected client-owned sessions #27 / Pylon release 0.0.9 PrimeIntellect-ai/prime-agent#84 because a replacement Pylon process cannot reconstruct the prior live-owner cleanup proof.
• Scope Windows to WSL2/Linux for the current parity program. Native Windows Prime execution and a native ACP fallback are deferred until upstream Prime supports Windows; authenticated native transport is not a current graduation gate.

Prime issue #20 and the separate Comet correlated-turn consumer are not dependencies of Pylon provider parity. Preserve that work independently; do not serialize Pylon delivery behind it.

Parity sequence

• Complete managed plan support for ACP behind a negotiated capability.
• Expose the missing Prime background-process lifecycle needed by Pylon.
• Add exact autonomous-run correlation and cancellation where Prime APIs are insufficient.
• Support heartbeat and scheduled-run projection after Pylon owns occurrence, turn, and checkpoint semantics.

For each item, search current Prime issues, pull requests, releases, and source first. Choose adopt, hybridize, retain, or redesign; do not remove Pylon behavior merely because upstream adds an overlapping feature.

[rynfar]

Coordination item opened: #3 defines the cross-repository correlated_prompt_lifecycle_v1 contract for exact daemon prompt ownership, event attribution, scoped cancellation, reconnect reconciliation, and usage isolation.

Implementation will use isolated branches after foundation PR #2 lands. The active foundation clone and the Pylon PR PrimeIntellect-ai#137 rebase remain untouched.

[rynfar]

Foundation status after PR #2:

• PR chore(pylon): establish fork governance and sync #2 merged into pylon as 703155b21bf53539c88bc19534fde448a9dff4bd.
• The first manual Pylon upstream sync run mirrored fork main to Prime upstream d60fab8a76d9c169f945341f0ee3bde21903bb55 exactly.
• Candidate integration found one real overlap conflict in packages/ai/test/prime-inference-models.test.ts and opened Prime upstream sync blocked #4, as designed. No candidate PR or branch was published.
• Issue feat(coding-agent): expose correlated prompt lifecycle to daemon clients #3 remains independently owned by its coordinating agent and is unblocked.

The sync workflow is functioning correctly; #4 now owns the required human overlap decision and resolution branch.

[rynfar]

Upstream integration landed:

• PR chore(pylon): merge Prime upstream through d60fab8a #5 merged into pylon as 12e1fdac74e120a1c1407d0c76d98bc18b73a07d after all hosted checks passed and two independent read-only reviews found no issues.
• The landed history contains exact Prime upstream d60fab8a76d9c169f945341f0ee3bde21903bb55 and exact Prime PR fix(coding-agent): kernel test fixtures speak protocol 3 PrimeIntellect-ai/prime-agent#1886 head a97f6995d39f2cccbd33123ab246f9786fc8ac13 as merge ancestry.
• The catalog overlap is recorded as hybridize; the upstream protocol-fixture repair is recorded as adopt.
• Prime upstream sync blocked #4 closed automatically and the integration branch was deleted.
• The one-review rule could not be satisfied because the repository has one collaborator who authored the PR; the maintainer explicitly approved the green head and it was merged with the admin bypass.

Remaining foundation work is limited to PR-token credentials, final branch/check policy, the inherited Linear workflow decision, and release naming/signing/provenance. Issue #3 remains owned by its coordinating agent.

[rynfar]

Fork operations checkpoint:

• pylon now requires strict GitHub Actions checks build-check-test and Check changelog fragment, conversation resolution, and PRs. Admin enforcement is on. Formal approvals are zero while rynfar is the only eligible reviewer; the count must return to one when a second reviewer exists.
• PR chore(pylon): disable inherited upstream workflows #6 removes the inherited Linear and release workflows from the default product branch and documents the reviewed workflow inventory and singleton protection. Both independent audits found no blocker; fresh hosted checks are running on 843891d30c61b67c08fa38b011461afaba1ffce0.
• An Actions-only main ruleset was rejected by GitHub because the built-in Actions integration cannot be a repository ruleset bypass actor. The safe fork-native fallback is viable: POST /repos/pylon-code/prime-agent/merge-upstream for branch main returned merge_type: none at the exact synced head. A follow-up will switch the sync script to that API before locking main with fork syncing enabled.
• Release audit confirms no releases, environments, repository secrets, or variables exist. Removing the inherited workflow is necessary but not sufficient; local publish commands, foreign npm package names, tag controls, and provenance remain release blockers.
• The optional custom PR token remains unset. Candidate-ready issues are the current fail-safe path; automatic draft PR credentials need a separate least-privilege decision.

Issue #3 remains untouched and independently owned.

[rynfar]

The focused fork-sync hardening is now in PR #7: #7

Reviewed head: fdf44c4acad2ba139c037f09357eb0ae852f26c6.

The change removes direct main pushes and the optional privileged-token path, gates GitHub fork synchronization behind an exact fail-closed repository ruleset, verifies exact post-sync mirror identity, and makes review candidates and their CI evidence deterministic. Two independent final reviews found no code blockers.

Rollout remains intentionally separate from the code change. After PR #7 lands and its workflow definition is trusted on pylon, maintainers must enable squash merging, provision and externally audit the zero-bypass exact-main ruleset, remove the superseded classic protection, then validate a real manual sync before calling main locked. Until provisioning, the new workflow will fail closed.

[rynfar]

The next parity candidate is now in coordinated unmerged PRs:

• correlated_prompt_lifecycle_v1: feat(daemon): add correlated prompt lifecycles #9
• Pylon integration and conservative stock-Prime fallback: fix(server): keep Prime background work out of foreground turns pylon#163

This hybridizes Prime's existing queue/background primitives behind an explicitly negotiated ownership, delivery, provenance, cancellation, usage, and recovery contract. The decision and upstream overlap are recorded in .pylon/features.yaml and .pylon/upstream-review.md.

This does not expose heartbeat creation or scheduled-run projection. Those remain separate until Pylon owns occurrence, turn, checkpoint, clear, stop, and deletion semantics. Both PRs remain unmerged pending hosted checks and maintainer approval.

[rynfar]

The coordinated lifecycle work has merged:

• Prime PR feat(daemon): add correlated prompt lifecycles #9 merged into pylon as 02a53ddf7d5557eb732a52e6d991b3d5d52430d1.
• Pylon PR feat(coding-agent): add slash command aliases PrimeIntellect-ai/prime-agent#163 merged into pylon as 486af3ea52180c3dcda51b6411ffaeecf9987d08.
• Both reviewed heads passed every hosted check; final independent audits found no remaining P0/P1 findings.

The optional correlated prompt contract and conservative stock-Prime admission fallback are now landed. Heartbeat creation, autonomous occurrence projection, and scheduled-run checkpoint ownership remain separate parity items; this tracker stays open.