Repository navigation
Description
Implementation progress - 1 November 2026
Partial implementation; acceptance remains open. Merged #269 supports explicit pause and resume for eligible queued deliveries before workflow start. The open queued-control follow-up #278 exposes that scope through the durable operation journal, CLI and MCP: malformed unaddressable commands are refused before journal insertion, known no-effect refusals are bound to the original command, and uncertain outcomes require receipt readback. A pending resume after restart requires explicit resubmission.
The follow-up also prepares an internal actor-bound ordinary publication handoff fence and staged-index checks. Public running-workflow pause remains disabled; the controlled handoff fixtures do not establish a general pause fence.
Verification: The unchanged reviewed source tree passed 66 affected local public, queued, operation, CLI/MCP, Temporal and publication cases. Final head 04cb084739c4c999869199bced40b3c3440bde41 on merged main 62f6d3877dfa8e22dc23cf78958f77443e6771e0 passed Check 41342543032 and current CodeQL. All five jobs, ZIP digests and four complete disjoint same-revision reports/JUnit were authenticated: 1,462 runtime passes and 188 source passes. Four existing runtime and three existing source opt-in/capability bodies remain skipped and unexecuted. This is configured hosted-gate qualification; provider/publication authority boundaries were controlled and full production acceptance remains open.
Publication: PR #278 is open and unmerged; its reviewed candidate and hosted evidence are ready for merge.
Remaining: Public running pause, coordinated children, in-flight effects, durable checkpoint/restart custody, and execution-time clock semantics remain unfinished. Native settlement depends on #242. Full stage-by-stage acceptance and absolute-deadline preservation remain required. This progress report does not close the issue; the required behavior and acceptance below remain authoritative.
Earlier references to PR #238 and foundation stack #253 as open describe historical candidates; both have merged.
Problem and current capability
The runtime persists execution history and supports cancellation plus specific stopped-run continuations. The inspected CLI/API control surface has no general durable pause/resume operation; a Project “paused” mapping is not an execution control.
The user must be able to request a pause at any point and return later with the work and state preserved. Pausing must not require cancelling and reconstructing a new delivery or keeping a chat open.
Required behavior
- Add authenticated, idempotent pause/resume operations for an individual delivery and its coordinated child work, exposed consistently through API, CLI, MCP and dashboard. Clearly state the scope before an explicit “pause all” action.
- Persist pause intent before acknowledgement. Report pause_requested/pausing/paused truthfully, including outstanding work/effects; accepting a request must not claim that processes have already stopped.
- Fence new role launches, test/gate starts, publications, merges and other owned effects once the pause request is accepted. Safely settle or cooperatively interrupt in-flight work and record any external operation that cannot be recalled.
- Reach “paused” only after owned native work is settled or safely checkpointed and remaining external uncertainty is explicitly represented. Provide a visible next action if safe settlement cannot complete.
- Persist the original authorization, accepted plan, candidate including dirty/unpublished work, sessions, checks/findings, effect journal, ownership and cumulative budgets using the durable preservation contract.
- Survive client disconnect, worker/API restart, host reboot and removal of a disposable chat worktree. No unattended resumption after those events.
- Resume the same logical work under explicit user instruction, revalidating current environment, ownership and external effects. Do not reset product budgets, broaden scope, create duplicate PRs or replay uncertain writes.
- Define clocks precisely: paused time does not consume execution-time allowance; an absolute user deadline remains absolute. Preserve historical timer semantics unless an explicit compatible amendment is made.
Acceptance
Pause during intake/planning, implementation, tests, review, CI wait and publication/merge readback; test duplicate pause/resume, pause-versus-completion races and restart between request persistence and acknowledgement. Include dirty/unpublished source, active child workers and an external write whose outcome is initially unknown.
For every scenario prove no new owned effect starts after the pause fence, state survives restart/cleanup, and explicit resume continues once from the correct checkpoint with original budget and authority. Cancellation remains a distinct operation.
Dependencies
Blocked by:
- feat(runtime): make long-running control operations durable and idempotent #236 — make long-running control operations durable and idempotent.
- fix(recovery): preserve resumable work independently of disposable worktrees #239 — preserve resumable work independently of disposable worktrees.
- fix(recovery): expose typed lifecycle transitions and actionable blockers #240 — expose typed lifecycle transitions and actionable blockers.
These are required contracts for this issue's acceptance. Related work without a blocking relationship can proceed in parallel.
Blocks:
- test(runtime): qualify recovery, pause and cleanup through public interfaces #246 — qualify recovery, pause and cleanup through public interfaces.
Activity
ebarti commentedon Oct 26, 2026
Implementation dependency update for stack #253: safe native pause/resume still depends on #239's authenticated writer cutoff and checkpoint custody, #243's exact reservation/GO accounting, and #242's actual descendant-settlement proof. These are genuine activation dependencies, not reasons to disable cleanup or accept a tracking projection as custody.
The local feature and timed controls remain gated. Real owned Temporal tests have exercised a denied native check followed by same-job continuation, and expiry receipt loss followed by exact SDK-history-bound readback. Those fixtures still control selected authority, snapshot/custody and tracker boundaries; they do not establish complete delivered-run diagnostics or all-child feature graph checkpointing. Independent child pause tokens must remain independent on parent resume.
The remaining implementation and #246 qualification must prove complete parent/child and diagnostic state transitions, exact original clocks and budget, no active source writer at checkpoint acknowledgement, and no duplicate external effects after crash, restart or lost response. This issue stays open; no installed runtime or active delivery has been changed.
ebarti commentedon Oct 27, 2026
Partial before-start support is proposed in #269. The exact candidate passed an owned managed-service test with a real Temporal server and worker across pause, restart and resume. This is limited to queued deliveries; pausing running workflows and complete native custody remain open, including the dependency on #242.
fix(runtime): preserve control lineage and publication custody (#257)
fix(api): preserve delivery operations across caller loss and restart (…
fix(delivery): bind authority to the original execution (#260)
fix(native): retain custody across the pre-ready launch window (#265)
feat(delivery): allow queued deliveries to start paused (#269)
fix(recovery): verify current checkpoint custody on readback (#275)