AGENTS.md

1# prompts/agents — Agent Contract
2 
3## Scope
4 
5This subtree owns role prompts for SDP agent personas and reviewers.
6 
7## Contract
8 
9Every top-level agent must state:
10 
11- what intent or workflow it supports
12- what SDP stage or entity it updates
13- what artifact, verdict, or handoff it must emit
14 
15## Dependencies
16 
17Agent prompts may reference root `AGENTS.md`, `docs/reference/agent-catalog.md`,
18and relevant skill files.
19 
20Do not copy full skill workflows into agent prompts. Agents choose and execute
21skills; skills own the procedural steps.
22 
23## Runtime Assumptions
24 
25Harnesses may load these prompts through symlinks, generated adapters, or explicit
26dispatch prompts. Keep role behavior harness-neutral unless the file is explicitly
27for one harness.
28 
29## Local Rules
30 
31- Prefer fewer durable roles over many overlapping personas.
32- If a role does not own a unique transition, collapse it into an existing agent
33  or a review dimension.
34- Review agents must emit findings with severity and evidence.
35- Execution agents must emit concrete changed files, gates run, and blockers.
36