Skip to content
Pocketto
Pipeline

Approval gates

Every point where the Pocketto pipeline stops for your approval, plus the independent review gates that stop for a reviewer who is not the author.

Pocketto distinguishes two kinds of gates. User authorization gates stop and wait for you. Independent review gates stop and wait for a reviewer that is not the author. You are not asked to approve its findings, but you are informed and consulted when recovery is exhausted.

User authorization gates

#GateWhereWhat your approval covers
1Spec approvalGrinding handoffThe spec's problem statement, scenarios, and constraints. Authorizes the handoff to planning, never implementation.
2Plan approvalPlanning handoffGeneration of derived execution artifacts (index.md, task files, phase manifests). Never implementation.
3Execution authorizationStructuring handoffStarting development. Asked once; a clear prior instruction covers it.
4Education choiceDevelopment startup in an Education repoWhether you implement with guidance or dispatch subagents.
5Closeout authorizationClosingAdvancing a reviewed phase and closing the plan. A passing verdict alone is not authorization.
6Push consentEnterprise phase publicationDurable (--push-consent phase-branches) or per-commit (consent-push --sha <40-hex>); never assumed from credentials.

Rules that protect these gates:

  • Approvals are not re-asked when an earlier instruction already covered them. You are never prompted twice for the same decision.
  • Silence is not approval. A consequential choice you have not made gets a question.
  • If you explicitly skip a gate or choose a simpler workflow, the agent honors it and reports which checks were skipped. Nothing claims a skipped gate passed.

Independent review gates (not user gates)

These run without you; their purpose is that no agent grades its own work.

ReviewStageReviewer
Advisor curationPitchingIndependent advisor (LLM-to-LLM)
Edge-case hunterGrinding, after GWT scenariosRead-only subagent
Spec reviewerPlanning, on the planRead-only subagent
Test-strategy auditPlanning, only when triggers fireRead-only subagent
In-loop auditDevelopment, per task after DONERead-only auditor (mechanical gate first)
Phase-level passDevelopment, after all tasks DONERead-only subagent

The main agent never judges code; implementers never self-report quality as evidence.

Bounded recovery

Every review gate has a bounded recovery contract: after the initial attempt, at most two materially different correction/review rounds. Reaching the cap is a mandatory user checkpoint (RECOVERY_CHECKPOINT), not a silent pass and not a BLOCKED verdict. It preserves the unresolved findings, the strategies tried, and one proposed next strategy, and waits for your explicit authorization to open another bounded window.

What does not burn a round: a clean pass, and auditor infrastructure failures (those get a bounded reroute: fresh auditor → advisor fallback → report the exact human action needed).

BLOCKED is reserved for a genuine human dependency: a decision, access, information, or authorization the agent cannot obtain. Every blocked report names the dependency, the evidence, and the concrete unblocking action, and safe independent work continues.

Example

Illustrative: the spec reviewer twice returns findings on the same task. The plan is corrected with a materially different strategy, re-reviewed, and findings remain. The skill stops at RECOVERY_CHECKPOINT, lists both rounds and the unresolved finding, and asks whether to authorize another bounded window. It neither hands off an unapproved plan nor silently drops the finding.

On this page