Capcut Video Editor Football Executor It Provides One Out

User outcome

A Work on the direct model API route whose calls include one with an unknown effect (a send, payment or other consequential call that may or may not have happened) now ends unknown, as it already does on the CLI route. The owner reads the unknown bubble that says what to check instead of an ordinary partial (or even success) closing line.

Runtime impact

One check moves: the Work-wide unknown-effect scan (_work_has_unknown_effect) now settles every attempt in _run_worker_attempts, right after the attempt returns, for both routes. The CLI route's own copy is removed; its behaviour is unchanged. No new state, store, tool or model call.

Goal / authority

Issue #663 (bug), requested directly by the owner; no delivery-plan.yaml change. No authority, Grant, approval or egress change. Non-goals: the read-side status predicates (#664), retry semantics (safe_retry already refuses replay from the effect records).

Constitution / specification alignment

• C6 / C8: the owner-facing record follows the recorded effect, not the worker's report; an unknown effect is never rewritten as completion.
• C16: no task-specific code; route parity for an existing invariant.
• Mega888 Deposits Withdrawals Trusted Toys TheThe Best Uk Ratings Music Many I1 (the unknown bubble is the effect owner's statement), Zus88 Fresh Game Topslot88id KomunitasBelow Is Update Date E1 (the one unknown rule), Live22 Slots New Elevate Slot MalaysiaLive22 Believes Being There Due Things closing lines (an unknown Work has no closing line).

What changed

• src/personal_agent/quickstart_service.py: _run_worker_attempts sets run.outcome='unknown', run.unknown_statement=run.response and run.resolved_blocker=False when _work_has_unknown_effect holds, after either route's attempt returns PROCEED and before the orchestration step reads the outcome. _run_cli_attempt drops its own copy of the same check.
• tests/test_unknown_outcome_routes_663.py: direct-route and CLI-route negative tests.

Interactions checked: _settle_worker_outcome's #752 upgrade (only partial/failed) and cli_shortfall (only succeeded) never touch unknown; orchestration_step already refuses re-delegation on an unknown reason and now also sees outcome='unknown' on the direct route; _steer_continuation_refusal returns unknown directly; _finish_work's unknown branch uses run.unknown_statement; the budget Continue offer reads only partial/failed. Parks for local authority or a connector still happen first on both routes, as before on the CLI.

Existing solutions review (Adopt / Adapt / Build)

• problem/boundary: route parity for the existing unknown-effect outcome invariant; the direct API route stored the worker's reported outcome instead of unknown.
• standard-library/platform candidates: none applies; this is Still's own Work/Evidence outcome record.
• official SDK/reference/standard candidates: none applies; no protocol or SDK behaviour is involved.
• mature open-source/framework candidates: none applies; the effect record is Still-owned C6/C8 state.
• internal repository candidates: AgentService._work_has_unknown_effect (the one Zus88 Fresh Game Topslot88id KomunitasBelow Is Update Date E1 rule) and the CLI route's existing assignment, adopted unchanged and moved to the single place both routes pass through.
• search evidence (paths/symbols/docs checked): quickstart_service.py_work_has_unknown_effect, _run_cli_attempt, _run_api_attempt, _run_worker_attempts, _settle_worker_outcome, _finish_work, _fail_work, orchestration_step, _steer_continuation_refusal, _offer_budget_continue; conversation_projection.terminal_text.
• maintenance/security/supply-chain fit: no dependency added.
• licence fit: no new code licence; AGPL-3.0-only unchanged.
• runtime/deployment/compatibility fit: same process, same stores; CLI route behaviour identical.
• decision: Adopt
• why rejected candidates are insufficient: a second copy of the check in _run_api_attempt would duplicate the rule; the shared loop placement keeps one copy and also lets the orchestration step and steer refusal read unknown on both routes.
• Still-owned policy/authority boundary kept outside the dependency: the outcome is derived from Still's own tool-event effect records; no worker report can override it.
• N/A reason: not applicable (decision is Adopt).

Authority / threat model

Data: the Work's own tool events and outcome row. No Grant, folder scope, network destination, secret or approval semantics change. Failure path closed: a direct-route worker that reports succeeded/partial/failed after a call whose effect is unknown no longer leaves an owner-facing record (status, cause, bubble, closing line) that differs from what ran. Abuse path: none new; retry and continuation stay refused from the effect records (safe_retry, may_continue), and an unknown Work is still never re-delegated.

AgentPackage / runtime delta

N/A

Supply-chain / trust evidence

N/A

Lifecycle / recovery impact

None beyond the outcome: an unknown direct-route Work is not retried or continued (unchanged), and now says so.

README / localization impact

N/A

Validation

• tests/test_unknown_outcome_routes_663.py (new):
  • direct route, run_agent reporting succeeded, partial and failed after an unknown-effect call: status unknown, cause and bubble are the unknown statement alone, safe_retry refuses, no budget Continue offer, goal summary effect unknown. Fails on the old code for all three outcomes (verified by reverting the source change).
  • direct route without an unknown call: the reported partial/failed outcome and its closing line are unchanged.
  • CLI route: an unknown call still ends unknown with the statement; without one it does not.
• Full suite locally (pytest -n auto --dist loadfile tests): 7003 passed, 74 skipped, 22555 subtests passed.

Evidence class / operating validation

Local deterministic, model-free tests (fake run_agent, fake CLI execution adapter, temporary stores). No live model or external service observed.

Evidence and completion audit

• Direct route ends unknown whatever run_agent returned: test_an_unknown_effect_call_ends_unknown_whatever_run_agent_reported.
• Bubble is the unknown statement: same test.
• Direct route without one unchanged: test_a_work_without_one_keeps_the_outcome_run_agent_reported.
• CLI unchanged: CliRouteUnknownOutcomeTests plus the existing CLI unknown tests (test_orchestrator, test_agency_recovery).
• Required CI on the exact head; Codex review of the final head (escalated).

Delegation

N/A

Escalated review (Codex)

• material security/authority boundary changed: yes
• trigger, if yes: the owner-facing record (status, cause, bubble) differed from what ran (AGENTS.md review guidelines, P1 class); authority, approvals and egress are unchanged.
• escalated review required: yes
• Codex review of the exact final head and each finding's disposition (fixed / answered), if required: pending, requested by the owner after CI.
• negative tests that hold the triggering boundary, if required: tests/test_unknown_outcome_routes_663.py (direct route for each reported outcome; direct and CLI controls).
• Codex reviewed the exact final head (asked once after remediation): pending
• post-review remediation changed the triggering boundary or introduced a new trigger (decides only an additional human or specialist re-review): N/A

AI assistance and accountability

• accountable human submitter: Jongtae
• AI tools / roles used (optional; any tool or none): Claude Code (implementer)
• self-review findings and remaining risks (a second model run by the author is self-review, not independent review): the direct route's turn-provenance status still records what the worker answered (answered/partial), as the CLI route records answered; the Work row carries unknown.
• this PR does not itself authorize a release: yes

Data and security impact

No new data stored or sent. The owner-facing outcome now matches the recorded effect on both routes.

Known limitations / follow-up

None known.

Historical authority / migration

Extends #598 I1 / #618 E1 to the direct route; earlier records are unchanged.

Closes#663

🤖 Generated with Claude Code