agent-acceptance

name	agent-acceptance
description	SADP governance orchestrator. Use when user says "@go", "@go read", "@go edit", "@go risky", or requests formal task execution with mandatory reviewer evidence. NOT for casual conversation.

agent-acceptance - SADP Governance Orchestrator

Role: orchestrator, not final judge. @go manages the workflow pipeline. The final pass requires deterministic guards, an independent reviewer artifact, and CI/human gates where applicable.

Modes

Command	Profile	Write Scope	Network	Use Case
@go read	read-only	none	off	Code review, analysis, audit
@go edit	ai-dev	project dir, deny secrets	off	Normal feature/bug work
@go risky	ai-risky	project dir, deny secrets	on	CI changes, deps, deploy config
@go	ai-dev	project dir, deny secrets	off	Standard formal task

Required State Machine

human_gate
  -> executor/fixer
  -> tester
  -> reviewer
  -> finalizer

The reviewer node is mandatory. A run cannot be marked passed without review.md and review.yaml created by a reviewer role that is separate from the executor/fixer.

Evidence Contract

Every @go run must create a run evidence directory containing:

File	Producer	Purpose
diff.patch	executor/fixer or harness	Real code delta
test-output.md	tester	Command output and exit codes
safety-report.json	deterministic guard	Secret/path/scope checks
chain-evidence.json	orchestrator/harness	Role/session/model chain
review.md	reviewer only	Human-readable review
review.yaml	reviewer only	Machine-readable review verdict
final-report.md	finalizer only	Deterministic summary

executor and fixer must not write review.md or review.yaml. finalizer must not substitute for reviewer judgment.

Reviewer Rules

The reviewer must:

• Run in a separate session/model identity from the executor/fixer.
• Read diff.patch, test-output.md, safety-report.json, and chain-evidence.json.
• Treat executor logs/reports as claims, not facts.
• Produce both review.md and review.yaml.
• Block if any P0/P1 finding is unresolved.

Minimum review.yaml:

reviewer_role: reviewer
reviewer_id: "<session-or-agent-id>"
executor_id: "<executor-session-or-agent-id>"
verdict: pass | blocked | fail | escalate
reviewed_inputs:
  - diff.patch
  - test-output.md
  - safety-report.json
  - chain-evidence.json
findings:
  - id: finding-001
    severity: P0 | P1 | P2 | P3
    status: open | resolved | false_positive
    title: "short finding"

Workflow

1. Gate 0: check AGENTS.md, rules, mode/profile, and TaskSpec allow_write.
2. Executor/fixer: dispatch with opencode run; initialize evidence with python tools/go_evidence.py init <run-dir> --run-id <id> --task .ai/tasks/<id>.yaml --executor-id <session-id>.
3. Tester: run commands; write test-output.md.
4. Guards: run python tools/go_evidence.py guard <run-dir> --task .ai/tasks/<id>.yaml; this captures safety-report.json.
5. Reviewer: dispatch a separate reviewer session; write review.md and review.yaml.
6. Finalizer: run python tools/go_evidence.py finalize <run-dir>.
7. Verdict: passed only if guard, reviewer, and evidence validation all pass.

Automated Dispatch

opencode run -m "deepseek/deepseek-v4-pro" -c "Read .ai/tasks/<id>.yaml, execute, write executor evidence to <run-dir>"
opencode run -m "deepseek/deepseek-v4-pro" -c "Review <run-dir>; write review.md and review.yaml; do not edit code"
opencode run -s <session_id> -c "continue instruction"

P0 Hard Stops

#	Rule
1	No destructive git without human approval
2	No secrets in code, logs, or reports
3	No command injection or path traversal
4	No fake green
5	No write outside TaskSpec allow_write
6	No pass without independent review.yaml
7	No pass with unresolved P0/P1 findings

Finalizer Rule

The finalizer is deterministic. It may summarize and validate artifact presence, but it must not make code-quality judgments or override the reviewer. If reviewer evidence is missing or invalid, final status is blocked.

Full protocol: docs/agent-runtime/sub-agent-dispatch-protocol.md Policy rules: .ai/policy.yaml enforced by tools/ai_guard.py