git:20260905.2ef932b to git:20260906.021b41c

6 added, 9 removed. Audit A to A.

---

name: implement

description: 'Execute one bounded RED to GREEN experiment from bead or caller intent; return derived subject identity and check facts. Triggers: "implement", "implement this bead", "run the experiment". Full plan-to-validation requests route to rpi.'

practices:

- tdd

- refactoring

- small-batch-flow

hexagonal_role: driving-adapter

consumes: []

produces:

- subject-manifest.v1

output_contract: 'subject-manifest.v1 digest, author context ID, and exact acceptance-check receipts returned through the response or runtime channel'

context_rel:

- kind: customer-of

with: plan

skill_api_version: 1

user-invocable: true

metadata:

graph_root: true

tier: execution

dependencies: []

capabilities: [execute_one_experiment, collect_factual_evidence]

effects: [modify_declared_subject, derive_subject_manifest]

canonical_status: canonical

disposition: keep

---

# Implement

Execute exactly one bounded experiment described by the resolved bead or caller

intent. Implement owns subject edits and factual evidence; the runtime derives

identity and receipts.

## Prompt

```text

Implement bead ag-1234 from its text: acceptance "ao gate check lists

skill.probe-coverage", scope cli/internal/gates/** plus regen outputs, first

check `cd cli && go test ./internal/gates/...`. RED first, smallest change,

return the manifest digest and check receipts, stop.

```

## It's working if

- The first transcript command is the acceptance check, and its output shows

the expected failure (or a green baseline for a relocation or refactor).

- Every path in `git diff --stat` falls inside the declared scope; an outside

consumer is reported as `file:line`, not absorbed.

- The response carries the `subject-manifest.v1` digest, author context ID,

and verbatim check output; no `git commit` or `git push` appears.

## Workflow

1. Read the intent, acceptance, and scope from their existing source; before

the first write, read `boundaries.md` in the rpi skill's `references`

directory for what Implement does not own.

2. Run the declared first acceptance check before changing behavior. RED-first

applies when acceptance is behavioral: preserve evidence that the check

fails for the expected missing behavior. Relocations, doc merges, and pure

refactors record an honest green pre-change baseline instead.

3. Make the smallest in-scope change that satisfies the active behavior.

4. Run the targeted acceptance checks and capture factual results.

5. Refactor only while those checks stay green. Refactoring does not change the

acceptance test.

6. Have the runtime derive actual changed paths and `subject-manifest.v1` from

the before/after subject.

- 7. When `scripts/evidence-orphans.sh` exists, have the runtime run it over the

- runtime-derived union of changed paths and append its output to the check

- receipts the validator reads, so orphaned evidence is a receipt like any

- other rather than a surprise at verify time. The receipt runs again after

- every repair round, over the union as it stands, because a repair can

- orphan evidence the first pass did not. Each entry carries a `cause`:

- `changed_path`, `digest_drift`, `both`, or `skill_changed`, which separates

- the evidence this change orphaned from drift that was already there. The

- runtime derives the orphan set; the model never lists it by hand.

+ 7. Run `bash scripts/evidence-orphans.sh <changed paths>` over the changed

+ paths the runtime derived, and put its output in the check receipts the

+ validator reads, so orphaned evidence arrives as a receipt rather than as a

+ surprise at verify time. Run it again after every repair round, over the

+ paths as they stand, because a repair can orphan evidence the first pass did

+ not. Read the output as written and never hand-list the orphans instead.

8. Return the manifest digest, author context ID, and exact check receipts in the

response or runtime channel. Stop.

Specialists (standards, domain, test, refactor, security) advise only. During

edits, run the smallest deterministic checks that can falsify the change,

reuse exact-input receipts whose subject and tool identity still match, and

run the full suite at the integration boundary unless the intent makes it the

first check.

## Scope conflict rule

On discovering a live consumer of the change outside the declared write scope

(a test asserting the old path, a generated twin, a gate reading the moved

file), stop and report the exact file and line to the caller, who may revise

the intent and start a separate invocation; a different acceptance contract

is a new intent.

Before declaring GREEN, self-audit the diff for mocks, placeholders, TODO

stubs, hardcoded fixture values, weakened assertions, regenerated goldens,

widened tolerances, suppression directives, or specification edits standing

in for real behavior. A changed test, gate, fixture, golden, or acceptance

source must be required by the original intent, with green coming from the

implemented behavior; a check that passes against a substitute or weakened

oracle is not evidence: finish the behavior or report it as not built.

## Boundary

Do not commit, push, claim, close, release, land, reserve, retry, or invoke a

semantic validator. A failed check is evidence for the caller, not permission

to create a packet or validation loop.