fix(skill): bicameral-report-bug — config.yaml keys-only by default (#200 A4)

CHANGELOG.md

@@ -40,6 +40,7 @@ All notable changes to bicameral-mcp are tracked here. Format loosely follows
 
 ### Fixed
 
+- `bicameral-report-bug` skill: `.bicameral/config.yaml` now defaults to keys-only inclusion in bug-report bodies (#200 finding A4). Workspace IDs, tokens, allowlists, and env-specific values are stripped by default; an explicit "include verbatim" toggle in the Step 3.5 transparency preview lets the operator opt in when the bug genuinely needs the values (e.g. a YAML parser regression). Defense-in-depth: the secret-redaction regex still runs on verbatim contents. Closes A4 in #200; A1 + A6 + partial A7 closed by #201.
 - Flow 3 e2e prompt clarified to use imperative shell phrasing for `git add` + `git commit` (#197). Resolves the "agent did NOT commit" flake observed on PR #194 + #195 e2e runs where the agent interpreted "stage and commit" as a non-shell verb. Adds a "Debugging Flow 3 fails" subsection to `tests/e2e/README.md` capturing the investigation order so the next maintainer has a starting checklist.
 
 ### Schema

plan-200-config-yaml-redaction.md

@@ -0,0 +1,86 @@
+# Plan: bicameral-report-bug — `.bicameral/config.yaml` keys-only by default (#200 A4) (example)
+
+**change_class**: feature
+
+**doc_tier**: minimal
+
+## Open Questions
+
+None. Scope and design are framed by #200's A4 finding and the user's directive: "transparency + accuracy + minimum data shared." PR #201 (just merged) closed A1 (python3 portability) and A6 (browser-open). Partial coverage of A7 (rationale-field drop). This plan addresses only the remaining A4 surface (`.bicameral/config.yaml` verbatim leak). (example)
+
+## Phase 1: Default config.yaml inclusion to keys-only with explicit verbatim toggle
+
+The current Step 2 instruction (post-#201) says: *"`.bicameral/config.yaml`: use `Read` on `.bicameral/config.yaml` if it exists. If `Read` errors (file missing), skip the section."* The Step 3 body assembly then dumps `<contents>` verbatim into the issue body. This violates the "minimum data shared" directive — config.yaml routinely contains team-server tokens, workspace IDs, allowlists, and environment-specific settings whose values aren't needed for diagnosing most bugs but whose presence is. (example)
+
+Phase 1 changes Step 2's instruction so the default included shape is the YAML's *key structure* (top-level keys, no values), with an explicit opt-in toggle exposed in Step 3.5's transparency preview that lets the user elect to include the verbatim contents when the bug genuinely needs them (e.g., a parser bug in config loading).
+
+### Affected Files
+
+- `skills/bicameral-report-bug/SKILL.md` — three edits, all on the existing skill markdown:
+  1. **Step 2 §"`.bicameral/config.yaml`"** instruction: replace the "use Read on .bicameral/config.yaml" line with the keys-only extraction rule (Read the file, parse the YAML or scan top-level non-indented keys, emit only the key list as a sorted bulleted block). (example)
+  2. **Step 3 body-assembly template**: replace the verbatim ```yaml <contents> ``` block in the assembled markdown with the keys-only block. Add a sentinel-line `(values redacted by default — opt in via Step 3.5 to include verbatim)`.
+  3. **Step 3.5 transparency preview**: add the `.bicameral/config.yaml` toggle to the operator-facing question. The toggle defaults to "keys only" and the operator can flip it to "include verbatim" if the bug needs the values. Update the redaction-summary block to print which mode was chosen (keys-only / verbatim) so the operator sees the choice in the preview. (example)
+
+### Changes
+
+**Step 2 §config.yaml** — replace the existing line with:
+
+```markdown
+- **`.bicameral/config.yaml`**: use `Read` on `.bicameral/config.yaml` if it exists. Extract ONLY the top-level key structure by default — every top-level YAML key (lines that don't start with whitespace), one per line, sorted alphabetically. Do NOT include values, nested keys, comments, or any other content. The keys-only shape is sufficient diagnostic signal for *"is this bug in the config loader?"* questions while leaking zero workspace IDs, tokens, or environment-specific settings. If the operator's bug genuinely needs the verbatim contents (e.g. a YAML parser regression), Step 3.5's transparency preview offers an explicit opt-in toggle. If `Read` errors (file missing), skip the section entirely. (example)
+```
+
+**Step 3 body-assembly template** — replace the existing ```yaml <contents> ``` block with the keys-only shape:
+
+```markdown
+  ## .bicameral/config.yaml   ← only if Read succeeded
+
+  ```
+  <sorted top-level key list, one per line, no values>
+  ```
+  *(values redacted by default — opt in via Step 3.5 transparency preview to include verbatim)*
+  ```
+```
+
+**Step 3.5 transparency preview** — extend the operator-facing question to include a config.yaml verbosity toggle. The exact `AskUserQuestion` shape:
+
+```python
+AskUserQuestion({
+  questions: [{
+    question: "Open the prefilled GitHub issue?",
+    header: "Open issue",
+    multiSelect: false,
+    options: [
+      { label: "Yes, open it (config.yaml: keys only)",
+        description: "Default — config.yaml top-level keys included, values redacted" },
+      { label: "Yes, but include config.yaml verbatim",
+        description: "Use only when the bug requires inspecting config values; values still pass the secret-redaction regex but workspace IDs / tokens / allowlists are exposed" },
+      { label: "Edit the body first",
+        description: "I want to revise the body in chat before opening" },
+      { label: "Cancel",
+        description: "Don't open anything; nothing leaves the machine" }
+    ]
+  }]
+})
+```
+
+The redaction-summary block printed before the question is updated to include the explicit current choice:
+
+```
+Auto-redacted in this body:
+  - ...existing redactions...
+  - .bicameral/config.yaml: keys only by default (toggle below to include verbatim)
+```
+
+When the operator picks "Yes, but include config.yaml verbatim", regenerate the body with the verbatim ```yaml <contents> ``` block (the existing pre-#200 shape, kept available for the opt-in path). Re-display the preview with verbatim contents and re-ask the open-issue question one more time so the operator sees what's actually being shipped before clicking through. The secret-redaction regex (`(api[_-]?key|token|secret|password|bearer)\s*[=:]\s*\S+`) still runs on verbatim contents — this is defense-in-depth, not a substitute for the keys-only default.
+
+### Unit Tests
+
+None for this phase. Per `doctrine-test-functionality` and the established precedent across plan-156 PR A Phase 2, plan-156b Phase 1, plan-187 Phase 2, plan-197 Phase 1: skill markdown is LLM-consumed agent-instruction, not pytest-invocable. The functional validation is the operator running `/bicameral-report-bug` against a real bug; correctness of the keys-only extraction, the toggle UX, and the regenerate-on-opt-in flow is observed at the runtime preview, not in unit tests.
+
+A static lint asserting the new instruction text contains the keys-only phrase would be presence-only by construction (it doesn't validate the LLM's behavior on the new instruction). Per the doctrine, presence-only assertions don't satisfy the test functionality bar.
+
+## CI Commands
+
+- `python scripts/lint_plan_grounding.py plan-200-config-yaml-redaction.md` — runs the plan-grounding lint shipped in PR #121 against this plan to catch any backtick-wrapped path tokens that don't resolve on the working tree.
+- `grep -n "<contents>" skills/bicameral-report-bug/SKILL.md` — sanity check that the verbatim-by-default `<contents>` placeholder is gone from the default body assembly. Expected after Phase 1: matches only inside the documented opt-in path (Step 3.5's "include verbatim" branch).
+- `grep -nE "config\.yaml.*verbatim|keys only" skills/bicameral-report-bug/SKILL.md` — sanity check the new instruction text and toggle copy are both present. Expected: ≥3 matches (Step 2 instruction, Step 3 template note, Step 3.5 toggle).

skills/bicameral-report-bug/SKILL.md

@@ -80,7 +80,17 @@ Sources (in order of preference, all read-only / non-bash):
   they leak business context. Record only the shape:
   `branch: <REDACTED>` and `recent commits: titles redacted`.
 - **`.bicameral/config.yaml`**: use `Read` on `.bicameral/config.yaml`
-  if it exists. If `Read` errors (file missing), skip the section.
+  if it exists. **Extract ONLY the top-level key structure by default** —
+  every YAML key whose line doesn't start with whitespace, one per
+  line, sorted alphabetically. Do NOT include values, nested keys,
+  comments, or any other content. Default-shape rationale: top-level
+  keys are sufficient diagnostic signal for *"is this bug in the
+  config loader?"* questions while leaking zero workspace IDs, tokens,
+  allowlists, or environment-specific settings. If the operator's bug
+  genuinely needs the verbatim contents (e.g. a YAML parser regression),
+  Step 3.5's transparency preview offers an explicit opt-in toggle —
+  see "Step 3.5 — Transparency preview" below. If `Read` errors (file
+  missing), skip the section entirely.
 
 Then assemble in your head (do NOT print to user yet):
 
@@ -119,9 +129,10 @@ Then assemble in your head (do NOT print to user yet):
 
   ## .bicameral/config.yaml   ← only if Read succeeded
 
-  ```yaml
-  <contents>
   ```
+  <sorted top-level key list, one per line, no values>
+  ```
+  *(values redacted by default — opt in via Step 3.5 transparency preview to include verbatim)*
   ```
 
 ---
@@ -186,6 +197,9 @@ Auto-redacted in this body:
     topic, intent, description, text, excerpt, title — parameter names kept,
     values replaced with <REDACTED>)
   - Branch name and commit subject lines in the Repo state block
+  - .bicameral/config.yaml: keys only by default (workspace IDs, tokens,
+    allowlists, env-specific values stripped — toggle below to include verbatim
+    if the bug requires inspecting config values)
   - API keys, bearer tokens, secrets, passwords (regex-matched)
   - <N> redaction(s) applied   ← print actual count, or "none detected"
 
@@ -209,8 +223,10 @@ AskUserQuestion({
     header: "Open issue",
     multiSelect: false,
     options: [
-      { label: "Yes, open it",
-        description: "Browser opens to a GitHub draft — you review and submit there" },
+      { label: "Yes, open it (config.yaml: keys only)",
+        description: "Default — config.yaml top-level keys included, values redacted" },
+      { label: "Yes, but include config.yaml verbatim",
+        description: "Use only when the bug requires inspecting config values (e.g. YAML parser regression). Values still pass the secret-redaction regex but workspace IDs / allowlists / env settings are exposed." },
       { label: "Edit the body first",
         description: "I want to revise the body in chat before opening" },
       { label: "Cancel",
@@ -220,7 +236,8 @@ AskUserQuestion({
 })
 ```
 
-- **Yes** → proceed to Step 4.
+- **Yes, open it (config.yaml: keys only)** → proceed to Step 4 with the keys-only body as already previewed.
+- **Yes, but include config.yaml verbatim** → regenerate the body, replacing the keys-only block in the `## .bicameral/config.yaml` section with the verbatim ```yaml <contents> ``` shape. Re-run the Auto-redacted summary on the new body (the secret-redaction regex still applies — defense-in-depth, not a substitute for the keys-only default). Re-display the transparency preview with the verbatim contents and re-ask the open-issue question one more time so the operator sees what's actually being shipped before clicking through. Then proceed to Step 4 once the operator confirms the verbatim shape.
 - **Edit the body first** → ask the user what to change, regenerate
   the body, return to this step.
 - **Cancel** → stop. Tell the user "Cancelled. Nothing was sent." and