Skip to content

Add capability-proved nonpersistent daemon workers #20

Open
#54
Open
Add capability-proved nonpersistent daemon workers#20
#54

Description

@rynfar
rynfar
opened on Aug 30, 2026

Outcome

Add a capability-gated fresh daemon-worker mode that does not create or append the worker recovery journal, and returns an exact create receipt proving that mode. This unblocks the host-private correlated Comet text turn without persisting its random lifecycle correlation token.

Coordination

An independent Comet #7 review found that the approved correlated lifecycle is copied into the worker recovery journal. The snapshot and correlated queued actions contain correlationId. Comet #7 explicitly forbids logging, serializing, persisting, hashing, or exporting its host-random correlation token. Comet cannot solve that below Prime's public API.

Public contract

Proposed names, to freeze during review:

  • frozen root SDK feature: nonpersistent_daemon_worker_v1;
  • daemon server capability: nonpersistent_daemon_worker_v1;
  • optional fresh-create field: workerRecovery: "disabled";
  • successful create summary receipt: workerRecovery: "disabled".

The mode is available only when all of these agree. Method presence, versions, schemas, an ignored create option, or the server offer alone are not proof.

The request is valid only for a fresh lifecycle:"client_owned" create with no sessionPath/resume selector. Ordinary creates omit the field and preserve current recovery behavior byte-for-byte. Older/stock artifacts ignore no new requirement and remain usable by consumers that do not request this mode.

Required behavior

  1. Validate the request at the supervisor before launch.
  2. Mark the worker descriptor as nonrecoverable without storing any prompt correlation or prompt payload there.
  3. Do not pass DAEMON_WORKER_RECOVERY_JOURNAL_ENV to that worker. No worker recovery journal may be created or appended during create, prompt ownership, queueing, delivery, terminal settlement, or cleanup.
  4. Never recover or reuse a nonpersistent worker after process loss. Preserve client-owned cleanup authority and retire its descriptor/process ownership safely.
  5. Return workerRecovery:"disabled" only for the exact worker generation launched in this mode. Contradictory, stale, resumed, or ordinary workers must not receive the receipt.
  6. Keep correlated lifecycle behavior, bounded ingress, attach proof, cancellation, and ordinary worker recovery unchanged.
  7. Classify the protocol change as backward-compatible and capability-gated. Increment the schema revision and update SDK/daemon docs plus Pylon ownership records.

Acceptance tests

  • SDK feature registry and server offer are exact and frozen/current.
  • A valid fresh client-owned create receives the exact disabled-recovery receipt.
  • Missing feature/offer, omitted option, malformed option, wrong lifecycle, saved-session selector, and ignored-option simulations cannot produce the receipt.
  • No recovery journal file exists before, during, or after a correlated prompt lifecycle whose correlation is a privacy canary; descriptor/session artifacts contain no canary.
  • A killed nonpersistent worker is not relaunched or recovered; authoritative client-owned cleanup still settles and removes ownership.
  • Ordinary resident/client-owned workers still write and recover from their journal under the existing tests.
  • Daemon restart/descriptor discovery treats a dead nonpersistent worker as nonrecoverable and never synthesizes a recovery lifecycle.
  • Exact-artifact no-model smoke binds the installed package to its package hash and proves the new create receipt plus current-generation correlated capability.

Merge order

  1. Implement and independently review this issue on a branch from current pylon.
  2. Produce and verify an exact package artifact.
  3. Update Comet chore(pylon): harden fork synchronization #7 to require the SDK feature, server offer, request option, exact create receipt, and existing post-attach correlated proof.
  4. Re-review Pylon's existing correlated consumer for compatibility; no provider behavior should change because it omits the new create field.
  5. Merge Comet chore(pylon): harden fork synchronization #7 only after both consumers approve the exact Prime artifact.

Activity

  1. rynfar commented on Aug 30, 2026

    @rynfar
    rynfar
    on Aug 30, 2026
    Author

    Claiming implementation.

    • repository: pylon-code/prime-agent
    • branch: feat/nonpersistent-daemon-worker
    • worktree: /Users/rynfar/.prime/worktrees/prime-nonpersistent-worker
    • base: pylon@91e13b6798343995291ccca6f523fba81ff96cd6
    • owned area: SDK feature registry/docs, daemon protocol/session-summary/worker-descriptor contracts, supervisor launch/recovery/receipt logic, focused daemon and issue-20 regressions, .pylon/features.yaml, and .pylon/upstream-review.md
    • shared contract: nonpersistent_daemon_worker_v1 plus fresh create/receipt workerRecovery:"disabled"; names remain subject to exact-head cross-consumer review
    • compatibility: ordinary creates and stock artifacts preserve current recovery; the new mode is optional, fresh, client-owned, non-resumable, and must never write a worker recovery journal
    • dependencies: Prime Expose attach-negotiated daemon session capability proof #17/PR feat(sdk): prove negotiated daemon session capabilities #18 are merged; Comet chore(pylon): harden fork synchronization #7 will remain blocked until an exact artifact from this issue is approved
    • validation: focused tests from the package root, npm run check, exact artifact/reproduction, journal privacy canaries, killed-worker cleanup/nonrecovery, and independent Prime/Pylon/Comet review
    • merge order: Prime Add capability-proved nonpersistent daemon workers #20 -> exact artifact -> Comet chore(pylon): harden fork synchronization #7 consumer repair -> compatibility re-review -> Comet PR

    I will not push or open a PR until a frozen exact head has no P0/P1 review findings.

  2. rynfar commented on Sep 9, 2026

    @rynfar
    rynfar
    on Sep 9, 2026
    Author

    Audit release and schema-31 replacement plan

    The cross-repository audit is complete. The earlier dirty checkout based on 91e13b6798343995291ccca6f523fba81ff96cd6 will remain untouched as a rejected reference. Prime origin/pylon is now 68603ed89bb597cd715fd6a77bc1c39d7e110298 (protocol 7, schema 30) and overlaps nearly every implementation seam through the newer caller-owned environment-cleanup and recoverable-owned-session adoption work. A direct rebase or cherry-pick would be unsafe.

    I am reclaiming #20 for a fresh transplant/redesign with this boundary:

    • repository/base: pylon-code/prime-agent from exact origin/pylon@68603ed89bb597cd715fd6a77bc1c39d7e110298;
    • branch: feat/nonpersistent-daemon-worker-schema31;
    • isolated worktree: /Users/rynfar/.prime/worktrees/prime-nonpersistent-worker-schema31;
    • rejected reference only: /Users/rynfar/.prime/worktrees/prime-nonpersistent-worker (must not be edited, staged, committed, reset, or cleaned);
    • owned source/docs/tests: .pylon/features.yaml, .pylon/upstream-review.md, packages/coding-agent/.changes/prime-20-nonpersistent-daemon-worker.md, packages/coding-agent/docs/sdk.md, daemon protocol/supervisor/worker/mode/CLI and command-journal sources, SDK feature definitions, and focused daemon protocol/client/supervisor/process/recovery tests needed by this issue;
    • shared contract: new schema 31; supervisor offer nonpersistent_daemon_worker_v1; strict fresh lifecycle:"client_owned" create with workerRecovery:"disabled"; exact same-generation create-only receipt; capability-negotiated fail-closed behavior; no recovery journal, recovery/adoption/relaunch/retry/residency/update snapshot, or persistent correlated identity for that worker;
    • current-contract integration: preserve schema-30 caller-owned environment cleanup and recoverable adoption for ordinary workers, but keep nonpersistent workers outside all recoverable-owned-session stores and adoption paths; bind expected recovery mode during worker authentication; fail descriptor downgrade closed;
    • compatibility: schema <=30 and ordinary schema-31 creates retain legacy recovery, replay, ownership, environment-cleanup, and adoption semantics; older clients never infer the optional mode; no Pylon provider source change is expected for this private process-lifecycle slice;
    • tests: deterministic schema/capability/request/receipt, full request identity, ordinary-vs-disabled reuse, journal privacy, aliases/retirement, attach fencing, descriptor downgrade, worker-auth mismatch, forbidden lifetime/adoption/update commands, environment-cleanup coexistence, real supervisor replacement, and full affected daemon/recovery/update suites;
    • dependencies: merged Expose attach-negotiated daemon session capability proof #17 / PR feat(sdk): prove negotiated daemon session capabilities #18 capability proof; Comet chore(pylon): harden fork synchronization #7 remains blocked on an independently reproduced exact Add capability-proved nonpersistent daemon workers #20 package artifact; Pylon umbrella pylon-code/pylon#114 stays cross-linked;
    • merge order: freeze and independently review exact Prime head -> reproduce and approve exact package artifact -> merge Prime Add capability-proved nonpersistent daemon workers #20 -> update/validate/review Comet chore(pylon): harden fork synchronization #7 against that artifact -> merge Comet chore(pylon): harden fork synchronization #7. No push or PR before all exact-head P0/P1 findings are resolved.

    I will preserve all schema-30 public behavior first, then lift only the still-needed safety invariants from the rejected checkout rather than transplanting the old diff mechanically.

  3. rynfar commented on Sep 10, 2026

    @rynfar
    rynfar
    on Sep 10, 2026
    Author

    PR #54 exact-head GitHub CI found a provider-neutral base-branch failure outside the nonpersistent daemon implementation.

    Build and check runs packages/ai build, which currently refetches live model catalogs. The live catalog has removed claude-sonnet-4.5 and other identifiers still used by the checked-in tests, so the generated union no longer compiles. The failure is unrelated to this feature and is reproducible from the base build path.

    Upstream has already merged the exact fix in PrimeIntellect-ai#2038 / commit 0894de1ded368f175b41baeb16ba3a58ec66f828: build the committed reviewed model catalog instead of refetching it. Its complete delta is one packages/ai/package.json script line plus its changes fragment.

    To make #54 independently green without weakening checks or committing unreviewed live generated data, I am extending this issue's owned files only to:

    • packages/ai/package.json
    • packages/ai/.changes/res-1269-deterministic-model-catalog-build.md

    I will import that exact upstream commit with provenance, then freeze a replacement head and repeat all five exact-head P0/P1 reviews, scrubbed CI, release tests, and two-build byte-for-byte artifact reproduction. No shared daemon contract changes. The imported upstream fix must merge with #54 before Comet #7 consumes the artifact.

    Failed job: https://github.com/pylon-code/prime-agent/actions/runs/34443252621/job/102762545810
    PR: #54

  4. rynfar commented on Sep 10, 2026

    @rynfar
    rynfar
    on Sep 10, 2026
    Author

    PR #54 process-smoke follow-up is now frozen at fc28e5604d104a0fafcda5f1766ec846caf2f363 (tree ad8dcdbda8bba834ff8b55d62314627e6856cc69).

    The prior Linux failure was a deterministic fixture bug, not a production transport failure: the nonpersistent process test supplied apiKey without a provider/model. Local Prime CLI auth selected a model and masked the omission; clean Linux had no model, so correlated admission failed safely with Nonpersistent correlated command failed.

    Proof and fix:

    • clean Docker Node 22 Linux at parent 0992c55…: focused regression reproduced the failure;
    • exact fc28e5604: the fixture writes an isolated custom model, selects it explicitly, and targets a reserved-then-closed numeric loopback endpoint;
    • no live catalog, host auth, DNS, or external inference endpoint is used;
    • Docker Linux fixed regression passed;
    • clean local focus passed 3/3;
    • full process suite passed 23 with 9 skips;
    • scrubbed full CI passed 341 files / 4,820 tests with 57 skips;
    • all five exact-head P0/P1 reviews approved;
    • two exact builds packed, verified, installed, and smoke-tested byte-identical artifacts.

    Validation receipt SHA-256: c3298680951edba9c72e79c4b975ed62f89f5849008a33c23d3b8adf7803ab84. PR #54 now points to this exact head and fresh GitHub checks are running.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions