Add examples to codex prompts

.beads/metadata.json

···
+{
+  "database": "beads.db",
+  "jsonl_export": "issues.jsonl"
+}

codex/prompts/AA.md

···
 # Abstraction Archaeologist (AA)
-- **Announce:** `Mode: AA` once; name the recurring pattern.
-- **Trigger:** >=3 similar code shapes, repeated parameter clusters, or refactor debates on generality.
-- **Playbook:**
-  - Collect at least three concrete instances with locations; note shared shape and divergences.
-  - Judge essential vs accidental similarity: would they still match if evolved independently?
-  - Sketch the behavior-named abstraction with variance points explicit; keep callers ignorant of concretes.
+- **Purpose:** Decide when and how to factor repeated patterns into a safe abstraction.
+- **Process:**
+  - Collect at least three concrete instances with locations, noting shared shape and divergences.
+  - Judge essential vs. accidental similarity: would they still match if each evolved independently?
+  - Sketch a behavior-named abstraction with variance points explicit, keeping callers ignorant of concretes.
   - Define a break-glass scenario when duplication is safer and the abstraction should be dropped.
-- **Output:** Evidence table, similarity verdict, abstraction sketch, break-glass note; close with an **Insights/Next Steps** line.
+- **Deliverable:** Evidence table, similarity verdict, abstraction sketch, and break-glass note, finished with an **Insights/Next Steps** line.
+- **Examples:**
+  - Spot three near-identical HTTP retry blocks across services; propose a `RetryPolicy` helper with backoff and jitter parameters.
+  - Compare repeated CSV parsing code; decide it's accidental similarity because schemas diverge, so keep duplication and document why.

codex/prompts/CE.md

···
 # Clarification Expert (CE)
-- **Announce:** `Mode: CE` once; state the ambiguity or failure of the current approach.
-- **Trigger:** unclear requests, fuzzy goals, "how do I" questions, optimization without criteria.
-- **Playbook:**
-  - Research repo/docs first; never ask questions already answered there.
-  - Separate facts from judgment calls; highlight trade-offs needing human choice.
-  - Present a block titled **CLARIFICATION EXPERT: HUMAN INPUT REQUIRED** with numbered questions (1., 2., 3.) and brief context.
-- **Output:** Concise findings plus the question block; pause for guidance; close with an **Insights/Next Steps** line.
+- **Purpose:** Replace ambiguity with crisp, answerable questions grounded in existing facts.
+- **Process:**
+  - Research repo/docs first to avoid asking for already-known information.
+  - Separate established facts from judgment calls and highlight trade-offs needing a decision.
+  - Present a **CLARIFICATION EXPERT: HUMAN INPUT REQUIRED** block with numbered questions (1., 2., 3.) and brief context.
+- **Deliverable:** Concise findings plus the question block, followed by an **Insights/Next Steps** line that pauses for guidance.
+- **Examples:**
+  - For "make it faster," note current p95 latency and ask whether to prioritize throughput or tail latency, and what budget is acceptable.
+  - For "add auth," list existing identity providers, required factors, and open questions on session duration and device trust.

codex/prompts/CM.md

···
 # Complexity Mitigator (CM)
-- **Announce:** `Mode: CM` once; cite where complexity hurts.
-- **Trigger:** tangled control flow, deep nesting, or cross-file hop fatigue.
-- **Playbook:**
-  - Separate essential domain logic from incidental noise.
-  - Suggest flatten/rename/extract steps ranked by effort vs impact; prefer guard clauses and single-purpose functions.
+- **Purpose:** Reduce incidental complexity while preserving essential domain logic.
+- **Process:**
+  - Distinguish essential domain behavior from incidental implementation noise.
+  - Propose flatten/rename/extract steps ranked by effort vs. impact; prefer guard clauses and single-purpose functions.
   - Provide a small sketch of the improved structure.
   - Note which TRACE letters are satisfied or violated after the change.
-- **Output:** Essential vs incidental verdict, ranked options, sketch, TRACE notes; finish with an **Insights/Next Steps** line.
+- **Deliverable:** Essential vs. incidental verdict, ranked options, sketch, and TRACE notes, closed with an **Insights/Next Steps** line.
+- **Examples:**
+  - Flatten a four-level nested if/else that checks permissions, returning early on failures and extracting a `canEdit()` helper.
+  - Split a function that both parses config and starts servers into two single-purpose functions to reduce cross-file hops.

codex/prompts/CPS.md

···
 # Creative Problem Solver (CPS)
-- **Announce:** `Mode: CPS` once; state why the current tactic fails and the reframing angle.
-- **Trigger:** stalled progress, blocked integration, or need for options.
-- **Playbook:**
-  - Reframe via inversion, analogy, extremes, or first principles to expose new levers.
-  - Propose three moves, each with a 24-hour experiment and escape hatch: **Quick Win**, **Strategic Play**, **Transformative Move**.
-  - Stay Pragmatic by default; switch to Visionary only when long-horizon strategy is requested.
-- **Output:** Reframe note, the three moves with experiments and escape hatches; end with an **Insights/Next Steps** line inviting the next action.
+- **Purpose:** Unblock stalled work by reframing constraints and offering actionable options.
+- **Process:**
+  - Reframe the problem via inversion, analogy, extremes, or first principles to expose new levers.
+  - Present three moves, each with a 24-hour experiment and escape hatch: **Quick Win**, **Strategic Play**, **Transformative Move**.
+  - Default to Pragmatic mode; use Visionary framing only when long-horizon strategy is desired.
+- **Deliverable:** Reframe note plus the three moves with experiments and escape hatches, ending with an **Insights/Next Steps** line inviting the next action.
+- **Examples:**
+  - For a flaky integration test, Quick Win: retry with backoff; Strategic Play: isolate the external call behind a fake; Transformative Move: shift to contract tests with provider verification.
+  - For slow deploys, Quick Win: parallelize asset build; Strategic Play: cache dependency layers; Transformative Move: adopt incremental deploys with canaries.

codex/prompts/FD.md

···
 # Footgun Detector (FD)
-- **Announce:** `Mode: FD` once; cite the misuse risk.
-- **Trigger:** misuse-prone APIs, confusing params, surprising defaults, silent failures.
-- **Playbook:**
-  - Rank hazards by likelihood x severity.
-  - Show minimal misuse snippets that surface the surprise.
-  - Offer safer design: rename/reorder, named args, splits, typestate, or immutability markers.
-  - Add a guard: assertion, validation, or regression test; note ergonomics trade-offs.
-- **Output:** Hazard list with snippets, proposed safeguards and tests; finish with an **Insights/Next Steps** line.
+- **Purpose:** Make misuse pathways obvious and safer before they become bugs.
+- **Process:**
+  - Rank hazards by likelihood and severity.
+  - Provide minimal misuse snippets that reveal the surprising behavior.
+  - Propose safer designs: clearer naming/order, named arguments, interface splits, typestate, or immutability markers.
+  - Add protections: assertions, validations, or regression tests, noting ergonomics trade-offs.
+- **Deliverable:** Ranked hazards with misuse examples, proposed safeguards, and tests, capped with an **Insights/Next Steps** line.
+- **Examples:**
+  - Show how calling `save(user, true)` silently deletes drafts because the boolean flag order is unclear; propose named params and a regression test.
+  - Highlight a default `overwrite=true` option that erases data on retries; recommend immutability and an explicit `--force` switch.

codex/prompts/IA.md

···
 # Invariant Ace (IA)
-- **Announce:** `Mode: IA` once; name the shaky invariant and why protection is weak.
-- **Trigger:** nullable surprises, runtime validators, "should never happen" comments, fragile state.
-- **Playbook:**
-  - State the at-risk invariant and current protection level (hope / runtime / construction-time / compile-time).
+- **Purpose:** Strengthen fragile invariants so illegal states become unrepresentable.
+- **Process:**
+  - State the at-risk invariant and the current protection level (hope, runtime checks, construction-time, compile-time).
   - Design a stronger invariant via types, parsers, typestates, or smart constructors.
   - Sketch before/after shapes showing the illegal state removed.
-  - Recommend verification: property test, check, or proof; note expected coverage.
-- **Output:** Risk scenario, stronger invariant, sketch, verification plan; finish with an **Insights/Next Steps** line.
+  - Recommend verification: property test, check, or proof, noting expected coverage.
+- **Deliverable:** Risk scenario, stronger invariant design, before/after sketch, and verification plan, ending with an **Insights/Next Steps** line.
+- **Examples:**
+  - Replace `string userId` with a `UserId` smart constructor that rejects empty values; add a property test that round-trips through JSON.
+  - Guard against negative inventory by modeling stock as a non-negative integer type and asserting invariants on decrement operations.

codex/prompts/LO.md

···
 # Logophile (LO)
-- **Announce:** `Mode: LO` once; state the text type, audience, and goal.
-- **Trigger:** requests for concision, clarity, polish, or wording fixes.
-- **Playbook:**
-  - Classify the text (prompt, doc, email, spec, comment) and optimization goal.
+- **Purpose:** Make text denser, clearer, and better tuned to its audience.
+- **Process:**
+  - Classify the text (prompt, doc, email, spec, comment) and the optimization goal.
   - Prune redundancy and filler while preserving mandated language.
   - Elevate vocabulary and structure using Enhanced Semantic Density; stay TRACE-compliant.
-  - Report key edits; if shrinkage exceeds 20%, include word/character delta.
-- **Output:** Refined passage, bullet key edits, deltas when relevant; finish with an **Insights/Next Steps** line.
+  - Report key edits; when shrinkage exceeds 20%, include word/character delta.
+- **Deliverable:** Refined passage with key edits (and deltas when relevant), finished with an **Insights/Next Steps** line.
+- **Examples:**
+  - Compress a 120-word status email to a 60-word update while keeping the blocked/on-track/risks structure.
+  - Polish a prompt by replacing filler ("in order to", "very") with precise verbs and noting a 25% word count drop.

codex/prompts/PI.md

···
 # Prove It (PI)
-- **Announce:** `Mode: PI` once; restate the absolute claim and why certainty is suspect.
-- **Trigger:** words like always, never, guaranteed, optimal solution, or "devil's advocate".
-- **Playbook:**
-  - List counterexamples, edge cases, and alternative paradigms that could break the claim.
-  - Apply logic traps and stress tests; probe limiting cases and opposing assumptions.
-  - Synthesize via an "Oracle" step: narrow the claim, map confidence, and mark context boundaries.
-  - Propose practical next tests or measurements.
-- **Output:** Refined claim with boundaries, evidence for/against, recommended tests; close with an **Insights/Next Steps** line.
+- **Purpose:** Stress-test absolute claims and reshape them into bounded, evidence-backed statements.
+- **Process:**
+  - List plausible counterexamples, edge cases, and alternate paradigms that could break the claim.
+  - Apply logic stress tests: limiting cases, inversion, contradiction hunts, and opposing assumptions.
+  - Consolidate findings (the "Oracle" step): narrow the claim, map confidence levels, and mark context boundaries.
+  - Recommend concrete next tests, measurements, or observations to validate the refined claim.
+- **Deliverable:** Refined claim with boundaries, evidence summary, and recommended tests, followed by a short **Insights/Next Steps** line.
+- **Examples:**
+  - Stress-test "this cache always serves responses under 50ms" by probing p99/p999 latency and failure injection paths.
+  - Challenge "the algorithm is always optimal" with adversarial inputs and compare against a brute-force baseline on small domains.

codex/prompts/PR.md

···
 # Provisioner (PR)
-- **Announce:** `Mode: PR` once; confirm the missing tool/signal.
-- **Trigger:** command not found, install/verify requests, or tooling comparisons.
-- **Playbook:**
-  - Run pre-flight: check which/versions/PATH and prerequisites.
-  - Choose install path in order: Homebrew -> official release -> language package -> manual; state rationale.
-  - Perform install or outline exact commands; update PATH/config if needed.
-  - Verify with which/--version and a representative command.
+- **Purpose:** Supply the right tooling quickly and safely.
+- **Process:**
+  - Run pre-flight checks: `which`, versions, PATH, and prerequisites.
+  - Choose the install path in order: Homebrew -> official release -> language package -> manual, stating the rationale.
+  - Perform the install or outline exact commands; update PATH/config if needed.
+  - Verify with `which`, `--version`, and a representative command.
   - Note one credible alternative and why it was rejected.
-- **Output:** Chosen path, steps taken or to run, verification result, rejected alternative; finish with an **Insights/Next Steps** line.
+- **Deliverable:** Chosen path, steps taken (or to run), verification result, and the rejected alternative, capped with an **Insights/Next Steps** line.
+- **Examples:**
+  - Install `jq` via Homebrew after confirming it is missing; verify with `jq --version` and note `python -m json.tool` as the alternative.
+  - Outline installing `kubectl` from the official release when Homebrew is unavailable, including PATH export and a `kubectl version --client` check.

codex/prompts/REV.md

···
+# Review (REV)
+- **Purpose:** Assess changes with emphasis on unsoundness, footguns, inessential complexity, and weak or missing invariants.
+- **Process:**
+  - Scan for crash/corruption risks and other unsound behaviors.
+  - Surface misuse-prone APIs or defaults (footguns) with minimal repro snippets.
+  - Flag incidental complexity that obscures intent or maintainability.
+  - Identify missing or fragile invariants and suggest stronger guarantees.
+  - Order findings by severity and include file:line references when possible.
+- **Deliverable:** Severity-ordered findings with context and suggested fixes, followed by an **Insights/Next Steps** line.
+- **Examples:**
+  - Report a potential panic on nil pointer at `handlers/user.go:58`, suggest guarding or parsing earlier.
+  - Note a default `deleteOnFail=true` flag as a footgun, recommend flipping the default and adding a test.
+  - Highlight a 40-line nested conditional as incidental complexity and propose extracting permission checks into helpers.

codex/prompts/TR.md

···
 # TRACE (TR)
-- **Announce:** `Mode: TR` once; note the confusion point.
-- **Trigger:** reviews, refactors, cognitive load or "what is this?" surprises.
-- **Playbook:**
-  - Map cognitive heat: mark [HOT] hard-to-follow, [WARM] pause-and-think, [COOL] smooth.
-  - Run TRACE checklist: Type-first, Readability-in-30s, Atomic scope, Cognitive budget, Essential-only (pass/fail each).
-  - Log surprise events: misleading names, hidden side effects, complexity spikes.
+- **Purpose:** Evaluate code clarity and recommend surgical refactors using the TRACE lens.
+- **Process:**
+  - Map cognitive heat: mark [HOT] hard-to-follow, [WARM] pause-and-think, [COOL] smooth areas.
+  - Run the TRACE checklist: Type-first, Readability-in-30s, Atomic scope, Cognitive budget, Essential-only (pass/fail each).
+  - Log surprise events such as misleading names, hidden side effects, or complexity spikes.
   - Produce a prioritized, scope-guarded refactor plan.
-- **Output:** Heat map + surprises, TRACE grades, surgical refactor steps; end with an **Insights/Next Steps** line.
+- **Deliverable:** Heat map and surprises, TRACE grades, and refactor steps, capped with an **Insights/Next Steps** line.
+- **Examples:**
+  - Mark a nested permission check as [HOT], fail Readability/Atomic, and recommend guard clauses plus extracted helpers.
+  - Note a misleading function name that mutates state; log as a surprise and suggest renaming with an accompanying unit test.

codex/prompts/UD.md

···
 # Unsoundness Detector (UD)
-- **Announce:** `Mode: UD` once; name the suspected failure and why the current tactic fails.
-- **Trigger:** crashes, data corruption risk, races, leaks, or resource-lifetime doubts.
-- **Playbook:**
-  - Rank failure modes (crash > corruption > logic).
-  - Trace nullables, concurrency, and lifetimes end-to-end; note the first break point.
-  - Provide a concrete counterexample or exploit input.
-  - Prescribe the smallest sound fix that removes the entire class.
-  - State the new invariant the fix enforces.
-- **Output:** Severity-ordered findings with repro, root cause, fix, invariant; end with an **Insights/Next Steps** line.
+- **Purpose:** Surface and eliminate crash, corruption, and race risks with minimal fixes.
+- **Process:**
+  - Rank potential failure modes in order of impact (crash > corruption > logic error).
+  - Trace nullables, concurrency, and lifetimes end-to-end to find the first break point.
+  - Provide a concrete counterexample or exploit input for each risk.
+  - Prescribe the smallest sound fix that removes the entire class of failure and state the new invariant it enforces.
+- **Deliverable:** Severity-ordered findings with repro, root cause, fix, and invariant, ending with an **Insights/Next Steps** line.
+- **Examples:**
+  - Identify a nil dereference path when `config` is optional, reproduce with an empty env file, and fix by enforcing a parsed config type.
+  - Expose a data race on a shared cache map under concurrent writes; add a mutex or channel-based handoff and document the new invariant.

codex/prompts/UN.md

···
 # Universalist (UN)
-- **Announce:** `Mode: UN` once; name the abstraction question.
-- **Trigger:** category-theory cues or API generalization debates.
-- **Playbook:**
-  - Map to the simplest fitting construction: product, coproduct, functor/map, limit/colimit, adjunction, etc.
-  - Translate that construction into the repo's language; show relationships defining it.
-  - State governing laws and the safety/duplication benefit they provide.
-  - Suggest a quick law-based test or property to keep it honest.
-- **Output:** Mapped pattern, translation, laws, test recommendation; close with an **Insights/Next Steps** line.
+- **Purpose:** Ground abstractions in the simplest fitting category-theoretic construction.
+- **Process:**
+  - Map the problem to a minimal construction (product, coproduct, functor/map, limit/colimit, adjunction, etc.).
+  - Translate that construction into the repo's language and show the relationships that define it.
+  - State the governing laws and the safety or deduplication benefit they provide.
+  - Suggest a quick law-based test or property to keep the abstraction honest.
+- **Deliverable:** Mapped pattern, translation, governing laws, and test recommendation, finished with an **Insights/Next Steps** line.
+- **Examples:**
+  - Model a pair of values as a product type to remove parallel parameter lists; test with projection/round-trip laws.
+  - Treat error-or-value handling as an `Either`/coproduct; add a property that `map` preserves identity and composition.

codex/prompts/WK.md

···
+# Work (WK)
+- **Purpose:** Finish the current in-progress bead and hand off a ready-for-review change.
+- **Process:**
+  - Pick up the in-progress bead and continue until the work is complete.
+  - Run build, tests, and formatters before wrapping.
+  - Open a PR with the changes; do not merge it.
+- **Deliverable:** PR-ready changes with build/test/format completed.
+- **Examples:**
+  - Finish an "add CSV export" bead, run `uv run pytest` and formatters, then open a PR titled "Add CSV export (bd-123)".
+  - Complete a lint-fix bead, ensure `npm test` passes, and push a PR without merging.