#871 Add capture-feedback skill

packages/agents/content/collections/recommended.md

@@ -3,6 +3,7 @@ name: recommended
 description: The batteries-included default set of artifacts, opt-in via codeassembly.yaml.
 dependencies:
   skills:
+    - capture-feedback
     - people-report
   subagents:
     - canary

packages/agents/content/skills/capture-feedback/SKILL.md

@@ -0,0 +1,64 @@
+---
+name: capture-feedback
+description: Capture feedback on agent behavior as a guidance-refinement candidate — apply the immediate fix when concrete, then record the generalized lesson. Use when correcting a misapplied rule or specifying a new desired behavior.
+user-invocable: true
+deploy: declared
+---
+
+# Capture feedback
+
+Record feedback on the agent's behavior as evidence for a future guidance refinement, applying the immediate fix when there is something concrete to change. The user describes a correction or a desired behavior, the agent applies it when applicable, and the agent appends a generalized record that a later recall-driven pass mines to refine guidance.
+
+The actual guidance refinement is **deferred** — this skill captures the candidate, it does not edit the skill, subagent, rulebook, general guidance, or helper. Refinement happens later, in bulk, via `kb-retrieve-events`, so a single data point never over-fits a rule.
+
+**Announce at start:** "Using capture-feedback to apply this and record it for guidance refinement."
+
+## When to use
+
+Invoke `/capture-feedback` whenever the user gives feedback that should change how the agent behaves — either:
+
+- **Correcting a misapplied rule** — the agent broke guidance that already exists (for example, title-cased a heading when the sentence-case rule was in force).
+- **Specifying a new desired behavior** — the user wants a behavior that no current guidance covers.
+
+Both are captured the same way; the difference is recorded in the tags.
+
+## Process
+
+### 1. Classify the feedback
+
+Decide one question: **did relevant guidance already exist?**
+
+- **Yes — misapplied existing guidance.** A rule was in force and the agent missed it. This is a mistake plus its correction.
+- **No — no such guidance.** The user is establishing a net-new expectation. There is no mistake, only a refinement to propose.
+
+Assess this best-effort from the guidance you can see. When you genuinely cannot tell, record the uncertainty in the body rather than forcing the call.
+
+### 2. Apply the immediate fix
+
+If the feedback names a concrete artifact — something the agent produced, such as code, a doc, a PR body, or a ticket — apply the fix:
+
+- **Apply it directly, without asking for approval.** Invoking this skill and describing the fix is the authorization.
+- **Confirm first only when the fix is genuinely ambiguous** (more than one reasonable reading of what to change) **or risky, large, or hard to reverse.** Then state the change you intend to make and wait for a go-ahead before applying.
+
+If the feedback is purely behavioral — a standing rule with nothing to fix right now — skip this step.
+
+### 3. Capture the record
+
+Invoke the `capture-event` skill to append an immutable record, composing its arguments and body as follows:
+
+- `--store codeassembly` — the project's agent-guidance KB. Pass a different `--store` only when the user directs the record elsewhere.
+- `--tags feedback` — always. Add `,mistake` when existing guidance was misapplied (step 1, "Yes").
+- `--skill <slug>` — when the refinement target is a skill.
+- `--summary` — a one-line recall label, for example "Agent title-cased a heading; sentence case is the rule."
+- **Body** — the generalized lesson, only to the extent needed to act on it later:
+  - The **error→correction pair** (misapplied-guidance mode) or the **desired behavior** (no-guidance mode), generalized — not the raw artifact or diff.
+  - A **best-effort candidate refinement target**: the guidance artifact and its type (skill, subagent, rulebook, general guidance, or helper). Use `--skill` for a skill target; name the artifact in the body for the other four. Write "candidate: undetermined" when you genuinely cannot place it.
+  - Any uncertainty about the classification from step 1.
+
+### 4. Report
+
+State what was fixed — or that the feedback was behavioral-only — and the captured record's id and path.
+
+## Completion
+
+The immediate fix is applied (when applicable) and an immutable `feedback`-tagged record is written, ready for a later refinement pass to mine via `kb-retrieve-events`.

packages/agents/content/skills/collaborate/SKILL.md

@@ -69,4 +69,4 @@ When you deem appropriate, proactively dispatch subagents to perform tasks. Good
 
 ## Skill improvement
 
-- When the user corrects a mistake caused by unclear skill definition, invoke the `capture-event` skill to record it: pass `--tags mistake` and `--skill <skill-at-fault>`, and put what went wrong and what the skill should have said in the body.
+- When the user corrects the agent, or specifies a new desired behavior, that feedback is evidence for refining a skill, subagent, rulebook, general guidance, or helper. Invoke the `capture-feedback` skill: it applies the immediate fix when there is something concrete and records a generalized `feedback` event — tagged `mistake` when existing guidance was missed — for a later refinement pass to mine.

packages/agents/src/lib/__tests__/capture-feedback-delivery.test.ts

@@ -0,0 +1,20 @@
+import { readFile } from 'node:fs/promises';
+import path from 'node:path';
+
+import { describe, expect, it } from 'vitest';
+
+import { resolveContentDir } from '../content-resolver.ts';
+import { resolveClosure } from '../dependency-resolver.ts';
+import { readDeploy } from '../deploy-frontmatter.ts';
+
+describe('capture-feedback delivery', () => {
+  it('ships capture-feedback as a declared skill in the recommended closure', async () => {
+    const contentDir = resolveContentDir();
+
+    const closure = await resolveClosure({ collection: ['recommended'] }, contentDir);
+    expect(closure.skills).toContain('capture-feedback');
+
+    const skillMd = await readFile(path.join(contentDir, 'skills', 'capture-feedback', 'SKILL.md'), 'utf8');
+    expect(readDeploy(skillMd)).toBe('declared');
+  });
+});