What fires when · The Harness Call-Map

Type skill agent hook CI workflow built-in cmd human gate · Trigger/effect AUTO BLOCKS ADVISES

local

① Intent

A human turns a story into a testable spec. No spec, no build.

spec-writer

skill · invoked

Writes specs/NNNN-name.md; runs the vague-line test; assigns the risk tier.

architect

agent · opus · invoked (HIGH)

System-design decisions that feed a HIGH spec — run before the build, not during.

eval-builder

skill · invoked (agentic)

Builds the versioned golden set next to the spec (§11).

Gate → Definition of Ready (human triage): acceptance checks pass the vague-line test, scope stated, risk tier set.

↓

local · inner loop

② Delegate

An agent plans, then builds inside the spec's bounds.

planner

agent · opus · suggested (M/HIGH)

Sequenced plan, human-approved before any code.

api-pattern

skill · invoked

Make the new surface match the codebase's canonical example.

test-writer

skill · invoked

Tests mapped to each acceptance check.

diagnose

skill · bug-fix specs

Reproduce with a failing test first, then fix, then prove.

debugger

agent · sonnet · invoked

Isolated root-cause; returns a distilled diagnosis + repro.

security-reviewer

agent · opus

Inline security pass as sensitive code is written.AUTO

build-error-resolver

agent · sonnet

Clears red builds with a minimal diff.AUTO on build fail

stop-gate

hook · every turn-end

Won't let the agent "finish" unless build + tests are green. AUTO BLOCKS

Inner loop: build → stop-gate blocks on red → build-error-resolver fixes → green → stop-gate lets go. "Done means the hook lets you stop."

↓

local

③ Discern — local (pre-push)

Prove it on the machine before it becomes a PR.

/code-review

built-in cmd · invoked

Diff-aware review; leaves a receipt.

/simplify

built-in cmd · invoked

Quality-only cleanup pass; leaves a receipt.

save-review-receipt

hook · auto

Records commit-bound review receipts.

review-gate

hook · on push / pr-create

No push until both receipts exist for HEAD. AUTO BLOCKS

pr-writer

skill · invoked

Branch → conventional commit → PR body mapping checks→evidence; spec in the diff; sets risk:high.

↓

CI · outer loop

③ Discern — CI (on the PR)

A fresh context re-checks the PR. Author ≠ grader is enforced by the clean checkout.

ci.yml

workflow · build-and-test

Build + test + coverage floor.BLOCKS

grader.yml

workflow · calls grader agent

Grades the spec file in the diff, line-anchored. ADVISES (required to run)

correctness.yml

workflow · fresh agent ≠ grader/author

Hunts changed lines for logic defects; accepted-risk:correctness overrides. BLOCKS on defect

security.yml

workflow · calls security-reviewer

Fires on gated paths or risk:high.BLOCKS on HIGH

eval-regression.yml

workflow · agentic changes only

When prompts/models/tools/behavior change (§11).BLOCKS on drop

Human Checker

human · non-author

Non-author approval; HIGH adds a named sign-off in the PR.BLOCKS

↓

④ Merge → Deploy

The merge bar clears; the proven artifact ships to dev.

merge bar

branch-protection · platform

CI green + grader ran + correctness passed + non-author approval (+ security + sign-off if HIGH). BLOCKS

deploy-dev.yml

workflow · on merge

Deploy the same artifact to dev; restore last good on failure.

/update-docs

built-in cmd · invoked

Sync docs to the shipped change. (No docs-drift workflow ships — this is a command, not a gate.)

eval-suite.yml

workflow · scheduled/manual

Periodic full benchmark — not per-PR.

The checking ladder = the spine (which piece attaches to each rung)

Rung	Check	What attaches here
1	Done-rule	`CLAUDE.md` — "done means the hook lets you stop" (persuasion only)
2	Agent re-checks each turn	Intrinsic agent behavior
3	Blocking Stop hook	`stop-gate` + `build-error-resolver` (local)
4	Separate grader	`/code-review` locally → `grader.yml` in CI (advisory)
5	Human / security gate	`correctness.yml` + `security.yml` + the human Checker (blocking)

Risk tier sets how far up you climb

Tier	What runs
LOW	Mechanical CI gates + grader advisory + a light human look. Ladder stops at rung 4.
MEDIUM	`planner` plan + grader + a non-author Checker. Rungs 1–5, standard permissions.
HIGH	All of it: `architect` at Intent, full ladder, `security.yml` pass, a named human sign-off, and `eval-regression.yml` if it's an agentic change. Tight agent permissions.