From harness
Generator ↔ Evaluator contract negotiation round. Generator proposes HOW, Evaluator reviews, iterate up to 3 rounds before code is written. Bridges the Planner's "what/why" to a testable "how".
How this command is triggered — by the user, by Claude, or both
Slash command
/harness:negotiateThe summary Claude sees in its command listing — used to decide when to auto-load this command
# `/harness:negotiate` — Generator ↔ Evaluator contract negotiation Runs AFTER [analyze.md](analyze.md) clears and BEFORE the Generator builds. Translates the Planner's high-level contract into a testable implementation plan that both Generator and Evaluator commit to. **When it runs automatically:** In the full [sprint.md](sprint.md) flow, `/harness:negotiate` runs automatically between ANALYZE and BUILD. You don't invoke it manually in the happy path. **When to invoke manually:** - `/harness:resume` landed on phase "negotiating" after session interruption - You edited the draft contrac...
/harness:negotiate — Generator ↔ Evaluator contract negotiationRuns AFTER analyze.md clears and BEFORE the Generator builds. Translates the Planner's high-level contract into a testable implementation plan that both Generator and Evaluator commit to.
When it runs automatically:
In the full sprint.md flow, /harness:negotiate runs automatically between ANALYZE and BUILD. You don't invoke it manually in the happy path.
When to invoke manually:
/harness:resume landed on phase "negotiating" after session interruptionThe Planner stays at the "what & why" level on purpose. File paths, component boundaries, data schemas, and API shapes are NOT in the architecture doc. This negotiation is where those details get pinned down — but by the agents that will actually build and test, not by the Planner upfront. This prevents the Planner from cascading wrong technical details into the entire downstream pipeline.
Generator proposes (writes proposal.md)
Evaluator reviews (writes review.md)
agreed / needs-revision) + specific asksIf needs-revision: Generator revises proposal (append new round to proposal.md), Evaluator re-reviews. Max 3 rounds.
If no agreement after 3 rounds: Escalate to human with:
On agreement: Generator writes final contract.md (overwrites Planner's draft) merging the original deliverables + negotiated implementation details + any Evaluator-added ACs.
Full shell-dispatch blocks (Round 1 Generator NEGOTIATE, Round 2 Evaluator REVIEW-PROPOSAL, final Generator FINALIZE-CONTRACT) live in sprint.md § 2c. NEGOTIATE. Reuse them verbatim.
review.md — if it's 3 lines, the Evaluator didn't actually engageGenerator and Evaluator MUST be separate subagents with fresh context. Do NOT self-evaluate.
npx claudepluginhub toolsbbb/belcort-harness --plugin harness