feat: pursue a goal autonomously #270
Conversation
…ng, and metadata reservation
…ta, and replay ignore
…nd behind goal-command flag
…ed by goal-command
…h budget threshold bands
…urnFlow afterStep
…with budget and step-cap stops
…st, flag docs, and gates
…plus loop-safety hardening
…int, not a fatal error
…ching + compaction)
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: e68378e8ba
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".
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: b0815f5ea1
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".
commit: |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: b7f34e153d
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".
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 734b1d438d
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".
| if (!stopHookContinuationUsed) { | ||
| const stopBlock = await this.agent.hooks?.triggerBlock('Stop', { | ||
| signal, |
There was a problem hiding this comment.
When an active goal hits its token budget, afterStep returns stopTurn, but runTurn still invokes this shouldContinueAfterStop hook before the goal driver can mark the goal blocked. In sessions with a Stop hook that returns a block/continuation, this branch appends the hook prompt and returns continue: true, causing another model step after the hard budget has already been reached. Check stopForGoalBudget before allowing steer/Stop-hook continuations so configured token ceilings remain deterministic.
Useful? React with 👍 / 👎.
| const reservedGoal = this.session.metadata.custom?.['goal']; | ||
| const patchCustom = (payload.metadata as Partial<SessionMeta> | undefined)?.custom; | ||
| if (patchCustom !== undefined && 'goal' in patchCustom) { |
There was a problem hiding this comment.
This guard only protects updateSessionMetadata, but the public create/fork session metadata path is copied directly into Session.metadata.custom, where goal is now reserved. An SDK caller that creates or forks a session with custom metadata { goal: ... } can still seed arbitrary goal state outside the lifecycle methods and flag gate, and later getGoal()/goal driving will treat that host metadata as the current goal. Apply the same reservation to create/fork metadata inputs.
Useful? React with 👍 / 👎.
| typeof state.status === 'string' && | ||
| ALL_GOAL_STATUSES.has(state.status) && | ||
| typeof state.turnsUsed === 'number' && | ||
| typeof state.tokensUsed === 'number' && |
There was a problem hiding this comment.
normalizeMetadata() relies on this predicate before treating metadata.custom.goal as a SessionGoalState, but this accepts records that lack required fields such as wallClockMs, createdAt, and actor timestamps. For an imported/forked session with the checked subset present, the malformed record is kept and later snapshots compute elapsed time from undefined, producing NaN/invalid goal status output instead of dropping the corrupt goal. Require the full numeric/string fields used by toSnapshot() and budget math here.
Useful? React with 👍 / 👎.
| await store.setBudgetLimits({ budgetLimits: budget, actor: 'model' }); | ||
| return { output: `Goal budget set: ${formatBudget(args.value, args.unit)}.` }; |
There was a problem hiding this comment.
When the model records a user-stated budget after the goal has already spent at least that much (for example it has used one turn and calls SetGoalBudget({ value: 1, unit: 'turns' }) as instructed), this persists an over-budget goal but returns an ordinary tool result. Because the goal driver only checks budgets at the turn boundary, the current tool batch and even the next model step can continue before the goal is blocked; return a stop hint or block immediately when the new snapshot is over budget.
Useful? React with 👍 / 👎.
chengluyu
changed the title
chengluyu
force-pushed
the
feat/goal-impl/1
branch
from
cf438a9 to
ecddc20
Compare
chengluyu
force-pushed
the
feat/goal-impl/1
branch
from
ecddc20 to
c0aacc0
Compare
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: c0aacc0f6e
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".
| if (choice === 'auto' || choice === 'yolo') { | ||
| if (!(await setPermissionForGoal(host, choice))) return; |
There was a problem hiding this comment.
In manual mode, if a goal already exists and the user enters /goal <new objective> without replace, choosing Auto/YOLO here changes the session permission before startGoal() later calls createGoal() and rejects with GOAL_ALREADY_EXISTS. The result is that no goal starts, but the session is left in a more permissive mode than before; preflight the goal creation conflict first or restore the previous permission when creation fails.
Useful? React with 👍 / 👎.
| // The wall-clock anchor is a runtime timestamp; a persisted one is stale | ||
| // (it predates the downtime). Drop it so resumed time isn't counted as | ||
| // pursuit — `resumeGoal` re-anchors a fresh interval. | ||
| state.wallClockResumedAt = undefined; |
There was a problem hiding this comment.
When a session is resumed with an active goal, clearing wallClockResumedAt here before demoting it to paused means applyStatus() cannot fold the elapsed active interval into wallClockMs. If the process exits or crashes while an active goal with a wall-clock budget has been running, that time is discarded on the next resume, so /goal resume can continue with more time remaining than the configured hard budget should allow.
Useful? React with 👍 / 👎.
| if (input.completionCriterion !== undefined && input.completionCriterion.trim().length > 0) { | ||
| state.completionCriterion = input.completionCriterion.trim(); |
There was a problem hiding this comment.
When the SDK or CreateGoal tool supplies a very large completionCriterion, this path trims it but bypasses the 4000-character cap applied to the objective, and the goal injector later repeats that criterion in every active/paused/blocked goal reminder. A long criterion can therefore bloat state.json and every continuation prompt despite the objective length guard; apply a comparable limit before storing it.
Useful? React with 👍 / 👎.
1 participant
This PR implements the “goal” feature, which has been highly requested for many times.
Related Issue
TBA.
Problem
TBA.
What changed
TBA.
Checklist
gen-changesetsskill, or this PR needs no changeset.gen-docsskill, or this PR needs no doc update.