Repository navigation
Checkpoint runner recovery, save task work and bound worker lifetimes - #692
Conversation
Codex Review Summary: ✅ CompletedThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 43321ee9f5
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
|
gatekeeper: Independent read of 11dc511, including the complete earlier 43321ee diff and this heads recovery-planning correction. Findings: no P1–P3 findings established in this checkpoint/save/deadline scope. Checked all changed code/tests/docs and surrounding task planning, trusted executable selection, inherited lock, sandbox supervisor and worker contracts. Repository revision reservation precedes task publication; observation/acknowledgement failures poison the writer and stop advancement, and pending observations recover before another allocation. Saved execution inputs precede mutable base/settings lookup, with a second revision check under the task lock. Worker-session capture is durable before continuation; worker completion and acceptance entry have separate revisions. Saves use ordinary sandboxed Git/hooks, preserve unresolved work and inspect cleanliness afterward. Status does not acquire flock, and inherited-descriptor ownership is considered after parent death. One saved deadline covers workers/repair/recovery; TERM sending cannot delay the final KILL. Tests exercise real CLI/bwrap, crash during commit, lock contention, frozen inputs, failed checkpoints and deadline work retention. This is a sizable addition, but remains sequential with bounded individual checkpoints and one bounded repository cursor rather than a daemon or per-task replay table. The observer is deliberately still a no-op; commit-bound publication evidence and object integrity are separately documented prerequisites, not claims of this PR. Current-head gate is green; author records the local process-limit limitation and preliminary footprint. I ran no builds and read no credential files. Touches oxa-work and sandboxed Git control: the orchestrator must still perform its own independent read after this comment. |
|
Independent security read at Carried forward (write these into the next slices):
🤖 Generated with Claude Code |
Tasks currently retain a branch but cannot reliably recover which worker/check/save action finished. This main-based replacement for #387 adds durable checkpoints, normal sandboxed saves, read-only status, and the decided task deadline (#333; #629 prerequisite). Tier T3: subprocess lock ownership and sandboxed Git callbacks.
Each transition has a repository-wide revision. A durable pending observation is replayed before any later allocation, with one bounded counter and a public CheckpointObserver hook; append failures stop the batch. Session IDs are saved immediately, repairs continue that session, and recovery uses the saved base, checks, offline policy and deadline. Unsupported formats are preserved and refused, following #331's current-format-only rule.
The runner saves successful, red, blocked and interrupted edits through ordinary Git hooks inside the existing sandbox. Read-only status never acquires a task lock, including when an orphaned worker/supervisor still owns the inherited descriptor. Task max_time defaults to 45 minutes; the child receives its remaining allowance through --max-time, with TERM at +60 seconds and KILL 10 seconds later. Interrupted work is kept.
The CLI observer is explicitly a no-op until the separately owned L1 store (#684) is integrated. This PR does not claim events/situation/standing commands are connected. Saved head/tree IDs describe saved work; commit-bound acceptance and object-integrity verification remain blocking prerequisites for publication in #397.
Validation:
See oxa/docs/work.md and oxa/docs/work-state.md for the recovery and observer contracts.