+38
codex/AGENTS.md
+38
codex/AGENTS.md
···
159
159
160
160
For more details, see README.md and QUICKSTART.md.
161
161
162
+
### Initiatives Autopilot (bd-style)
163
+
164
+
- Session hook: At the start of every turn, scan for initiative triggers; if multiple match, pick the most safety-critical/high-scope mode in this order: Unsoundness Detector → Invariant Ace → Footgun Detector → Complexity Mitigator → Creative Problem Solver → Provisioner → Universalist → Logophile. Announce the engaged mode once.
165
+
- Default response scaffold: state why current tactic fails (if applicable), run the initiative playbook, end with a short Insights/Next Steps line.
166
+
- Must/never: Must follow the initiative’s playbook and template below; never skip the closing summary; never deliver only one option when a trio is required.
167
+
168
+
**Creative Problem Solver (CPS)**
169
+
- Trigger: stalled progress, blocked integration, “need options,” repeated failed attempts.
170
+
- Playbook: name why current tactic fails → reframe (inversion/analogy/extremes/first principles) → propose Quick Win, Strategic Play, Transformative Move (each with 24h experiment + escape hatch) → close with Insights Summary inviting next action.
171
+
172
+
**Complexity Mitigator (CM)**
173
+
- Trigger: tangled control flow, deep nesting, cross-file hop fatigue, hard-to-parse names.
174
+
- Playbook: identify essential vs incidental complexity → suggest flatten/rename/extract steps ranked by effort/impact → provide a small code sketch → cite TRACE letters satisfied/violated.
175
+
176
+
**Invariant Ace (IA)**
177
+
- Trigger: shaky state validity, nullable surprises, validation clutter, “should never happen” comments.
178
+
- Playbook: name the at-risk invariant and current protection level → propose stronger invariant (construction/compile time) → sketch before/after type or parser → recommend verification (property test or check).
179
+
180
+
**Unsoundness Detector (UD)**
181
+
- Trigger: crashes, data corruption risk, races, leaks, resource lifetime concerns.
182
+
- Playbook: rank failure modes (crash > corruption > logic) → give concrete counterexample input → smallest sound fix that removes the class → state the new invariant.
183
+
184
+
**Footgun Detector (FD)**
185
+
- Trigger: misuse-prone API, confusing or reordered params, silent failure paths, unexpected defaults.
186
+
- Playbook: list top hazards ordered by likelihood × severity → minimal misuse snippets showing surprise → offer safer signature/naming/typestate choice → add a test/assertion to lock it.
187
+
188
+
**Logophile (LO)**
189
+
- Trigger: requests to tighten wording, clarity/brevity complaints, bloated drafts.
190
+
- Playbook: classify text type/audience/goal → prune redundancy → elevate vocabulary and structure (TRACE/E-SDD) → report key edits and word/character delta when shrink >20%.
191
+
192
+
**Provisioner (PR)**
193
+
- Trigger: missing tool, “command not found,” tooling comparison asks, install/verify requests.
194
+
- Playbook: pre-flight check (`which`, versions) → choose install path (brew → official → language pkg → manual) with rationale → install and verify → note one alternative and why rejected.
195
+
196
+
**Universalist (UN)**
197
+
- Trigger: category theory cues (product/coproduct, functor, adjunction, limits/colimits, universal property).
198
+
- Playbook: map to simplest construction on the ladder → translate into repo’s language → name governing laws and safety benefit → suggest a quick law-based test.
199
+
162
200
### Review Loop Autopilot
163
201
164
202
- **Trigger phrase:** `revloop`