Best fit
A long-running coding agent can declare a task complete, pause, compact, or hand off while required rows, dependencies, evidence checks, or recoverable work still remain.
// Workflow Automation
Agent Task Completion Gate
Verify an AI coding task against required rows, dependencies, evidence, artifacts, and real external boundaries before the agent declares completion or hands off.
What problem does Agent Task Completion Gate solve?
A long-running coding agent can declare a task complete, pause, compact, or hand off while required rows, dependencies, evidence checks, or recoverable work still remain.
Use it to
- Prevent an autonomous coding agent from ending while required queue rows remain
- Audit a new-session handoff against durable project state after context compaction
- Separate genuine account or publication boundaries from ordinary unfinished work
- Verify checksum-pinned artifacts and dependencies before closing a release task
What it returns
- Machine-readable COMPLETE, REVIEW, BLOCK, or CANNOT_ASSESS result
- Completion ledger for every required item and dependency
- Severity-ranked evidence, artifact, boundary, and claim-conflict findings
- Exact remaining-work handoff plus SHA-256 receipt
A representative input and result
input Objective: publish and verify ten catalog rows. Claimed status: COMPLETE. Rows 1-9: COMPLETE with ledger evidence. Row 10: PENDING. Required ledger artifact: present with matching SHA-256.
result Status: REVIEW
Finding: REQUIRED_WORK_REMAINS for row-10.
Finding: CLAIM_CONFLICT because claimed COMPLETE conflicts with the evidence-derived REVIEW state.
Remaining work: publish and verify row-10.
Artifact: ledger exists and checksum matches.
Access and approval boundaries
- Read access to the user-supplied completion packet and named local artifacts.
- Write access only to the selected local output directory.
- No browser, account, credential, environment variable, external service, or network access is required.
Known limitations
- The result is limited to rows, dependencies, evidence references, and artifacts represented in the supplied packet.
- A named URL, product ID, or external receipt is recorded but not independently fetched or authenticated by the offline helper.
- COMPLETE confirms that the supplied closure contract passed; it does not prove unrepresented external state or authorize another action.
- A poor or incomplete task decomposition can omit work that should have been required.
// choose with context
Is Agent Task Completion Gate the right skill?
It returns
Machine-readable COMPLETE, REVIEW, BLOCK, or CANNOT_ASSESS result Completion ledger for every required item and dependency
Do not use it as
It does not perform an account, payment, publication, identity, legal, correspondence, or representational action identified by the packet.
Compatibility, access, version, and licence
- Reads or accesses
- Declared local scope: agent-completion-input/**, agent-completion-output/**, *.json, *.csv.
- Compatibility
- Codex, Claude Code, and other agents that can read a normalized local JSON packet. Repository-backed plans, ledgers, handoff files, task queues, and local artifacts. Python 3 for the included standard-library completion helper.
- Version
- 1.0.0
- Licence
- Review the package licence and seller terms at the linked destination.
Questions
Does it keep an agent running automatically?
No. It produces a deterministic closure decision and exact remaining-work handoff. The surrounding agent or operator decides how to continue.
What counts as a genuine blocker?
A blocker must name a concrete owner decision, account, credential, pricing, publication, payment, identity, legal, correspondence, representation, CAPTCHA, or externally controlled state, plus a reason and evidence reference.
Can elapsed time or task difficulty return BLOCK?
No. Those are not qualified external boundaries. The helper returns REVIEW so safe required work remains visible.
Does it prevent duplicate marketplace listings?
It can preserve product IDs, URLs, and ambiguous rows as evidence and force REVIEW when they conflict, but it does not access or change the marketplace itself.
Does COMPLETE prove the live system is correct?
No. COMPLETE means the supplied closure packet passed. Any external state not represented and independently verified remains outside the result.
Does it need network or account access?
No. Version 1 runs locally with Python 3 and uses only the standard library.
Agent Task Completion Gate keeps proof and approval boundaries visible.
The listing includes the tested package, realistic samples, declared permissions, and known limitations.
Get Agent Task Completion Gate on Agensi