@event4u/agent-config 2.20.1 → 2.23.0

package/.agent-src/commands/agent-status.md

@@ -57,6 +57,22 @@ Extract from latest record:
 Pricing source: [`bench/pricing.yaml`](../../bench/pricing.yaml). Reader
 implementation: [`scripts/cost/track.mjs`](../../scripts/cost/track.mjs).
 
+### 3b. Read caveman delta + per-conversation cost lens
+
+Run two read-only Python helpers (stdlib-only, no-op safe if JSONL missing):
+
+- `python3 scripts/caveman_stats.py --format json` — per-session +
+  per-conversation + lifetime caveman delta. Honors suspended
+  multiplier (see [`docs/contracts/caveman-telemetry.md`](../docs/contracts/caveman-telemetry.md)) — delta reads `0` while suspended; display version + ACTIVE/SUSPENDED state regardless.
+- `python3 scripts/cost_by_conversation.py --format json` — per-conversation
+  total cost + model breakdown for current conversation, sourced
+  from same `agents/cost-tracking/sessions.jsonl` ledger.
+
+Surface in dashboard as one line:
+`[caveman: {lifetime.delta_tokens:+,} tok lifetime · {current_conv.delta_tokens:+,} this conv · multiplier v{multiplier_version} {ACTIVE|SUSPENDED}] · [conv cost: ${current_conv.total_cost_usd:.4f}]`.
+
+If both JSONLs missing or empty, omit line silently.
+
 ### 4. Calculate freshness thresholds
 
 - **Message threshold**: Next multiple of 25 ≥ current count

package/.agent-src/rules/caveman-speak.md

@@ -56,6 +56,8 @@ Post-rewrite validator runs on every reply when `speak_scope != off`:
 The rule documents the algorithm; agents apply it inline before
 sending. The mechanism is the rule, not a hidden script.
 
+Optional CI-side regression lock: [`scripts/validate_caveman_carveouts.py`](../../scripts/validate_caveman_carveouts.py) takes pre/post reply pair and asserts byte-identical preservation across all seven carve-out categories — runtime mechanism stays algorithmic; script is offline check.
+
 ## Caveman grammar
 
 - Drop articles (`the`, `a`, `an`).

package/.agent-src/skills/adversarial-review/SKILL.md

@@ -1,6 +1,6 @@
 ---
 name: adversarial-review
-description: "ONLY when user explicitly requests adversarial review, devil's advocate analysis, stress-testing a plan, or 'poke holes in this' — NOT for regular code review or design feedback."
+description: "ONLY when user requests adversarial review, devil's advocate, stress-test, OR honest critique of finished work ('poke holes', 'be brutal', 'was hältst du davon') — NOT for routine code/design review."
 personas:
   - critical-challenger
 source: package
@@ -16,6 +16,7 @@ Use this skill when:
 - You've completed a plan, design, or proposed fix and are about to present it.
 - The change is non-trivial (affects multiple files, changes behavior, touches critical paths).
 - You're about to recommend an architecture or design decision.
+- The user submits **finished work** (draft, post, naming decision, design proposal) and asks for an honest critical take — "what do you actually think?", "be brutal", "was hältst du wirklich davon". The flow is the same Attack-Defend-Revise loop, but on the user's artifact rather than the agent's plan.
 
 Do NOT use when:
 - The task is trivial (renaming, formatting, simple config change).

package/.agent-src/skills/canvas-design/SKILL.md

@@ -64,11 +64,12 @@ Document it in `philosophy.md` under `## Subtle reference`.
 Produce `agents/design-assets/{slug}/{slug}.{pdf|png}`:
 
 1. Pick the execution tool (Pillow, matplotlib, SVG, or framework-native)
-2. Limited palette — 2–5 colors, intentional and cohesive
-3. Geometric or organic forms per philosophy
-4. Text — sparse, design-forward, integrated as visual element; never overlapping, never falling off canvas
-5. Margins — every element contained, breathing room
-6. Repeating patterns, layered elements, systematic markers as the philosophy permits
+2. **Font selection** — pick a font that earns the philosophy. Inspect a working dir (e.g. `agents/design-assets/{slug}/fonts/` — create if missing) and place the chosen file there before render. System defaults (Arial, Helvetica, DejaVu, the matplotlib default sans) are the AI-template tell; reach for them only as deliberate fallback, never as the unexamined default
+3. Limited palette — 2–5 colors, intentional and cohesive
+4. Geometric or organic forms per philosophy
+5. Text — sparse, design-forward, integrated as visual element; never overlapping, never falling off canvas
+6. Margins — every element contained, breathing room
+7. Repeating patterns, layered elements, systematic markers as the philosophy permits
 
 ### 5. Refinement pass
 
@@ -103,10 +104,14 @@ If the user requests a series, treat each page as a story beat — distinct but
 * **No artist mimicry** — copying a living artist's signature style is copyright risk and breaks the original-work mandate. Propose an original direction.
 * **Text discipline** — most pieces fail because text creeps in as paragraphs. Words are visual accents, not explanation.
 * **One canvas** — single page unless multi-page is explicitly requested.
-* **Font availability** — the environment may not ship your target font. Pick a fallback before render time, or download into the working dir first.
+* **Font availability** — the environment may not ship your target font. Procedure step 4.2 governs selection; pick a fallback before render time and place the file in the working fonts dir.
 * **Output location** — always `agents/design-assets/{slug}/`. Never write binary artifacts to the repo root or to source-of-truth dirs.
 * **Refinement loop is real** — first render is the draft, not the deliverable.
 
+## Craftsmanship standard
+
+The deliverable is judged against human-crafted work, not against AI-generated comparables. Visible deliberation, intentional asymmetry, and palette restraint are the markers; default templates, generic gradients, and centered safe layouts are the failure mode. If the artifact would pass for "auto-generated stock visual", it has not earned its place — refine until intent is legible.
+
 ## Frugality Standards
 
 Apply the [Frugality Charter](../../contexts/contracts/frugality-charter.md).

package/.agent-src/skills/compress-memory/SKILL.md

@@ -0,0 +1,119 @@
+---
+name: compress-memory
+description: "Use when shrinking always-loaded memory files (AGENTS.md, CLAUDE.md, .cursorrules) via caveman grammar — refuses sensitive paths, round-trips via .original.md backup."
+source: package
+domain: process
+execution:
+  type: assisted
+  handler: internal
+  allowed_tools: [Bash]
+---
+
+# compress-memory
+
+> **Experimental.** Output-side caveman dialect did not meet kill-criterion in [`bench/reports/caveman-v1.md`](../../../bench/reports/caveman-v1.md) (`vs_terse` median −9.27 %). Input-side memory compression is orthogonal use case: savings target always-loaded memory budget, not reply stream. Treat ship-criterion as **per-target measurement**, not v1 verdict.
+
+## When to use
+
+Use when:
+
+- Always-loaded memory file (`AGENTS.md`, `CLAUDE.md`, `.cursorrules`, `GEMINI.md`, `.windsurfrules`) close to or above host tool's char budget and maintainer wants to recover input-token headroom.
+- Consumer-shipped `templates/AGENTS.md` failing `agents-md-thin-root` cap and pointer-extraction options exhausted.
+- Maintainer asks to "compress this memory file" or "shrink AGENTS.md" or names input-side caveman.
+
+## Do NOT
+
+- Compress reply, commit message, PR body, ticket summary, or any deliverable written *for* human reader — those are carve-outs in [`caveman-speak § Carve-outs`](../../rules/caveman-speak.md) and stay verbatim.
+- Compress path matching sensitive-file denylist (`.env*`, `.netrc`, `credentials*`, `secrets*`, `id_rsa*`, `*.pem|key|p12|pfx|crt|cer|jks`, `.ssh/*`) — script refuses with `SensitivePathError` and so should you.
+- Compress generated file (`.agent-src/`, `.augment/`, `.claude/`, `.cursor/`, `.clinerules/`, `.windsurfrules`) — edit source in `.agent-src.uncompressed/` and regenerate via package's sync + generate-tools scripts (`scripts/compress.sh --sync` + `scripts/compress.py --generate-tools`).
+- Hand-edit compressed memory file in place — run `--decompress` first; next compress pass refuses on body-hash drift (`CompressionRefused`).
+- Commit compressed file without committing matching `.original.md` backup — round-trip breaks otherwise.
+
+## Procedure
+
+1. **Analyse target first.** Before any write, **inspect** target with `view` or `wc -l` to confirm it is always-loaded memory file (`AGENTS.md`, `CLAUDE.md`, `.cursorrules`, `GEMINI.md`, `.windsurfrules`), not generated, and has prose paragraphs to compress (pointer-only Thin-Root file may net near-zero). Skip rest of procedure if any check fails.
+2. **Check denylist gate.** Run `python3 scripts/compress_memory.py <path> --check` — exit 0 = safe; exit 2 = denylist hit, stop and surface refusal.
+3. **Record baseline.** `wc -c <path>` — capture pre-compression char count for commit message.
+4. **Compress.** `python3 scripts/compress_memory.py <path>`. Script writes `<path>.original.md` (verbatim backup) and rewrites `<path>` with `original_sha256:` + `compressed_at:` frontmatter.
+5. **Inspect diff.** Eyeball every Iron-Law fence, numbered-options block, code fence, backtick span, `❌`/`⚠️`/`✅` line, and frontmatter pair — all must be byte-identical. Body prose may have lost articles (`the`/`a`/`an`) and auxiliaries (`is`/`are`/`was`/`be`/`that`/`which`).
+6. **Validate idempotency.** Re-run `python3 scripts/compress_memory.py <path>` — clean re-run is no-op (body hash matches). Non-zero exit = stop, escalate.
+7. **Commit both files together.** `<path>` and `<path>.original.md` ship as pair. Backup is rollback path; never commit one without other.
+8. **Rollback path.** If readability fails review at step 5: `python3 scripts/compress_memory.py <path> --decompress` restores backup and deletes `.original.md`.
+
+## Output format
+
+Maintainer-facing report after invoking script MUST contain, in this order:
+
+1. **Diff line** — pre/post `wc -c` as single line (`AGENTS.md: 2,891 → 2,453 chars (−15.1 %)`).
+2. **Backup path** — full path of `.original.md` backup so maintainer can verify it landed on disk.
+3. **Carve-out check** — one line confirming seven carve-out classes round-tripped (`carve-outs: 7 classes preserved · idempotent re-run: clean`).
+4. **Exit-code surface** — on failure, surface verbatim exit code and exception name (`SensitivePathError → exit 2`, `CompressionRefused → exit 3`, `FileNotFoundError → exit 4`); do not paraphrase.
+
+Do **not** narrate algorithm, grammar rules, or carve-out theory — rule and this skill document contract; output reports result.
+
+## Carve-outs — byte-for-byte preserved
+
+Mirrors seven carve-out classes in [`caveman-speak`](../../rules/caveman-speak.md). Compression engine in [`scripts/compress_memory.py`](../../../scripts/compress_memory.py) preserves:
+
+1. **Triple-backtick fences** — any language, any depth.
+2. **Numbered-options lines** — `^>?\s*\d+\.\s` plus `**Recommendation:**` / `**Empfehlung:**` label.
+3. **Backtick spans** — file paths, command names, identifiers inside body prose.
+4. **Status / error markers** — lines starting with `❌`, `⚠️`, `✅`.
+5. **Iron-Law ALL-CAPS lines** — `^[A-Z][A-Z0-9 ,.\-_/']{3,}$`.
+6. **Frontmatter blocks** — `---` fence pairs at head of file.
+7. **Mode markers** per [`role-mode-adherence`](../../rules/role-mode-adherence.md).
+
+Mangling any of these breaks Iron-Law surface host tool reads. Unit tests in `tests/test_compress_memory.py` lock each carve-out class as regression case.
+
+## Idempotency contract — Step 9 guard
+
+Script is **idempotent on clean re-runs**: running it twice on same target is no-op because body hash matches recompressed hash. Script **refuses** on **body drift**:
+
+| State | Outcome |
+|---|---|
+| No frontmatter SHA marker | Compress + write backup + inject SHA. |
+| SHA marker present, body re-compresses to same hash | No-op (return target unchanged). |
+| SHA marker present, body hash diverged | **Refuse** with `CompressionRefused` exit 3. |
+
+If you need to edit compressed memory file, run `--decompress` first, edit restored `.original.md` content, then re-run compressor. Never hand-edit compressed body — next CI run will either silently corrupt your edit (if it happens to re-compress to same shape) or hard-fail next compress pass.
+
+## Sensitive-path gate
+
+Every read path passes through [`scripts/validate_safe_paths.py`](../../../scripts/validate_safe_paths.py) `assert_safe()` before bytes leave disk. Gate is security floor for Phase 2 (input-side compression) per `step-16-caveman-substance.md` Phase 0; rollback of gate is rollback of this skill.
+
+CLI exit codes:
+
+- `0` — compress / decompress / check succeeded.
+- `2` — `SensitivePathError` (path matched denylist).
+- `3` — `CompressionRefused` (body hash diverged from frontmatter SHA).
+- `4` — `FileNotFoundError` (no `.original.md` backup to restore).
+
+## Gotchas
+
+- **Body-hash drift after manual edit** — hand-editing compressed body breaks `original_sha256:` invariant. Next compress pass refuses with `CompressionRefused` (exit 3). Recovery: `--decompress`, edit restored body, re-compress.
+- **`.original.md` backup missing on `--decompress`** — exit 4 (`FileNotFoundError`). Either someone deleted backup or `--decompress` already ran. Restore from git history; never regenerate backup by hand (regenerated content would not be byte-identical).
+- **Denylist false positive** — sensitive-looking filename outside denylist surface (project-specific naming) will still pass `assert_safe()`. Denylist necessary but not sufficient; maintainer responsible for never feeding secrets to compressor.
+- **Frontmatter ordering with existing keys** — if target already has frontmatter, compressor preserves existing keys, drops any prior `original_sha256:` / `compressed_at:` entries, and appends new pair. Other agents reading file should treat SHA + timestamp pair as canonical compression marker, not file size.
+- **Negative savings on pointer-heavy files** — `templates/AGENTS.md` already following Thin-Root (≥ 40 % pointers, ≥ 60-char *why*-clauses) has little prose left to drop; compression may net near-zero or even add bytes via frontmatter. Run [`agents-md-thin-root`](../agents-md-thin-root/SKILL.md) first to maximise pointer share, then measure whether this skill still pays.
+- **Generated-tree drift** — compressing `.agent-src.uncompressed/templates/AGENTS.md` does NOT propagate to `.augment/`, `.claude/`, etc. until package's sync + generate-tools scripts run (`scripts/compress.sh --sync` + `scripts/compress.py --generate-tools`). Always regenerate after compressing templated file.
+
+## Measurement — when to compress
+
+No published `caveman-v2` baseline for input-side savings yet (Step 11 of `step-16-caveman-substance.md` ships that). Until then, maintainer judges per-target whether compression pays its readability cost. Suggested workflow:
+
+1. `wc -c <path>` before — record baseline char count.
+2. `python3 scripts/compress_memory.py <path>` — compress + back up.
+3. `wc -c <path>` after — record post-compression char count.
+4. Eyeball diff: does prose stay legible? Are all Iron-Law fences intact?
+5. If yes → commit both `<path>` and `<path>.original.md`. If no → `--decompress`.
+
+Future `caveman-v2.md` will tabulate realised input-token saving against `agents-md-thin-root` 40 % pointer-ratio constraint so maintainer has numerical floor.
+
+## Cross-references
+
+- [`caveman-speak`](../../rules/caveman-speak.md) — runtime rule script mirrors for input-side targets; `caveman.speak_scope` does **not** gate this script (input-side runs regardless).
+- [`scripts/validate_safe_paths.py`](../../../scripts/validate_safe_paths.py) — Phase 0 gate; ported from upstream Caveman `63a91ec`.
+- [`scripts/compress_memory.py`](../../../scripts/compress_memory.py) — implementation.
+- [`tests/test_compress_memory.py`](../../../tests/test_compress_memory.py) — regression locks for each carve-out + idempotency + denylist.
+- [`docs/contracts/compression-default-kill-criterion.md`](../../../docs/contracts/compression-default-kill-criterion.md) — v1 verdict (output-side; informs but does not gate this skill).
+- [`agents-md-thin-root`](../agents-md-thin-root/SKILL.md) — caps consumer-shipped `templates/AGENTS.md`; this skill is one tool to land under cap.

package/.agent-src/skills/fe-design/SKILL.md

@@ -199,6 +199,14 @@ Step indicator (1 — 2 — 3)
 5. **Loading states** — Skeleton screens or spinners, never blank screens
 6. **Error recovery** — Clear error messages with suggested actions
 
+## Aesthetic direction
+
+Audit-pinned tokens and components always take precedence (see `existing-ui-audit`). When the audit pins an aesthetic, honor it without deviation. When the audit shows **no pinned aesthetic** — greenfield surface, marketing landing page, brand-new feature without design-system precedent — the design brief is allowed (and expected) to commit to a deliberate direction instead of defaulting to safe centered hero + 3-column features + CTA.
+
+Pick one direction up front and let composition, typography, and color follow from it. Avoid the "neutral AI default": uniform grid, system fonts as the visible body face, purple-to-blue gradients on white, predictable spacing. A direction that fits the brand intent (editorial / brutalist / refined / playful / retro / maximal / minimal / etc.) and is consistent across the page beats hedging.
+
+Surface the chosen direction in the design brief as a one-line statement (e.g. `aesthetic: editorial-magazine — asymmetric grid, serif display + sans body, generous gutters`). The apply step (`react-shadcn-ui` / `blade-ui` / `livewire` / `flux`) reads this line and matches typography, spacing, and motion to it; if no line is present, the apply step uses project defaults.
+
 ## Procedure
 
 When `directives/ui/design.py` (or any caller) cites this skill:

package/.agent-src/skills/prompt-optimizer/SKILL.md

@@ -29,14 +29,30 @@ domain: product
 
 1. **Deconstruct** — extract core intent, key entities, output shape, constraints; map what's provided vs missing.
 2. **Diagnose** — audit clarity gaps, ambiguity, missing specificity, missing structure; flag unstated assumptions.
-3. **Develop** — pick techniques by request type:
-   - *Creative* → multi-perspective + tone anchoring
-   - *Technical* → constraint-based + precision focus
-   - *Educational* → few-shot examples + clear structure
-   - *Complex* → chain-of-thought + systematic framing
+3. **Develop** — pick technique + template by request type:
+   - *Creative* → multi-perspective + tone anchoring (template: **CO-STAR** or **CRISPE**)
+   - *Technical* → constraint-based + precision focus (template: **RTF** or **File-Scope**)
+   - *Educational* → few-shot examples + clear structure (template: **Few-Shot** or **RISEN**)
+   - *Complex / multi-step* → chain-of-thought + systematic framing (template: **CoT** or **ReAct**)
+   - *Image AI (Midjourney / SD / DALL·E)* → **Visual Descriptor** or **Reference-Image-Edit**
    - Assign an AI role/expertise; layer context; add logical structure.
+   - Full template catalogue + when-to-pick rubric: [`docs/guidelines/prompt-templates.md`](../../../docs/guidelines/prompt-templates.md).
 4. **Deliver** — output the optimized prompt + a short "what changed" + (DETAIL only) techniques applied + one pro-tip.
 
+## Setting awareness
+
+The skill reads `prompt_optimization.outbound` (or `.default` when no
+outbound override is set) from `.agent-project-settings.yml`:
+
+| Mode | Behaviour |
+|---|---|
+| `off` | The skill refuses; the dispatcher echoes the user's prompt verbatim with a one-line note. |
+| `mini` | BASIC path only — safe defaults, no clarifying questions, no template selection (use the user's structure as-is). Hard cap: 1 turn. |
+| `max` *(default)* | Full 4-D + template selection. DETAIL mode auto-detects per the table below. |
+
+Any prompt starting with the configured `prompt_optimization.bypass_prefix`
+(default `/raw`) is echoed verbatim, no shaping, no template.
+
 ## Modes — BASIC vs DETAIL
 
 **Auto-detect on first turn:**
@@ -100,6 +116,7 @@ Format per § Output format. Do **not** execute the optimized prompt yourself un
 - The model tends to over-engineer BASIC mode — for a one-line ask, the optimized prompt should still be short. No 800-word system prompts for "help with my resume".
 - Don't drift into German welcome text. The optimized prompt mirrors the user's source-language preference; the skill's own scaffolding stays English (per `language-and-tone` for `.md`).
 - The model tends to **mix languages** in the optimized prompt when the user wrote in German but named an English-speaking target audience — pick one language for the whole optimized prompt body (default: source-language of the rough prompt unless the user explicitly named the target audience's language).
+- The model tends to inherit upstream dogma that "only 5 techniques are safe" (few-shot, role, structured-output, constraint-based, chain-of-thought). That claim travels with `nidhinjs/prompt-master` and is **rejected here** — CO-STAR, RISEN, CRISPE, ReAct, and the image-AI templates land in [`docs/guidelines/prompt-templates.md`](../../../docs/guidelines/prompt-templates.md) and are first-class. Pick by request type, not by upstream whitelist.
 
 ## Do NOT
 
@@ -107,3 +124,10 @@ Format per § Output format. Do **not** execute the optimized prompt yourself un
 - Do NOT ask more than one clarifying question per turn (`ask-when-uncertain` Iron Law).
 - Do NOT add an "I'm Lyra" preamble on every turn — the welcome belongs to the command entry point, not every reply.
 - Do NOT modify project files — this skill is conversational, no file writes, no commits.
+- Do NOT restructure a prompt that starts with the configured `bypass_prefix` (default `/raw`). Echo it verbatim with a one-line note.
+
+## See also
+
+- [`refine-prompt`](../refine-prompt/SKILL.md) — engine-inbound sibling; same `prompt_optimization` setting controls its mode
+- [`docs/guidelines/prompt-templates.md`](../../../docs/guidelines/prompt-templates.md) — 12-template catalogue cited from Develop step
+- AI Council session: `agents/council-responses/prompt-master-mini.json` (2026-05-17) — analysis behind template adoption and the 5-safe-dogma rejection <!-- council-ref-allowed: ADR decision trace -->

package/.agent-src/skills/react-shadcn-ui/SKILL.md

@@ -49,6 +49,15 @@ Do NOT use when:
 - Every interactive primitive must declare a focus-visible state via
   `focus-visible:ring-2 focus-visible:ring-ring`; that comes for free with
   the generated primitives but is easy to remove during a refactor.
+- **Anti-AI-slop: shadcn-default look.** The out-of-the-box shadcn
+  theme + `Inter`-as-system-fallback + neutral grays reads as
+  template across projects. Unless `state.ui_audit.design_tokens`
+  pins the neutral palette as the project's identity, the polish
+  step should match typography and color tokens to the design
+  brief's `aesthetic:` line (from `fe-design` aesthetic-direction).
+  Theme/font drift within a single audited project breaks
+  consistency — variation lives between projects, not between
+  components in the same surface.
 
 ## Covered primitives
 

package/.agent-src/skills/refine-prompt/SKILL.md

@@ -54,6 +54,50 @@ calling command (`/work`) owns prompt capture; this skill only refines.
 If `raw` is missing, empty, or whitespace-only the resolver already
 raised `PromptResolverError`. The skill never receives that input.
 
+## Modes and bypass
+
+The skill honours `prompt_optimization.inbound` (or
+`prompt_optimization.default` when no inbound override is set) from
+`.agent-project-settings.yml` / `.agent-settings.yml`. Three modes:
+
+| Mode | Behaviour |
+|---|---|
+| `off` | The skill is a no-op. The dispatcher writes `confidence={"band":"high","score":1.0}` directly and the engine proceeds with the literal prompt. No assumption inference, no clarifying questions. |
+| `mini` | Stack-aware light shaping. Steps 1-2 run; step 3 only emits `assumes:` lines for *implicit stack constraints* (framework, package manager) detected from config files. Steps 4-5 produce 3 AC bullets max. Low-band halts ask at most one question; medium-band halts are auto-confirmed silently. |
+| `max` *(default)* | Full procedure — every step 1–6 runs. Medium-band halts surface the assumption list verbatim; low-band halts ask one clarifying question. This is the existing behaviour. |
+
+**Bypass prefix.** If the raw prompt starts with the configured
+`prompt_optimization.bypass_prefix` (default `/raw`), the skill
+becomes a no-op regardless of mode. The dispatcher strips the
+prefix, passes the remainder through verbatim, and records
+`bypass:true` in the envelope so downstream surfaces (delivery
+report, `--no-prose-synthesis`) can attribute the skip.
+
+```
+/raw migrate auth.service.ts to use jose, keep the API shape
+```
+
+`/raw` is reserved at the prompt boundary only — it has no meaning
+mid-prompt and is not stripped when it appears inside the body.
+
+### Stack-config read (mini / max only)
+
+When the mode is `mini` or `max`, step 3 may read these config files
+(read-only, scope-locked) to enrich the `assumes:` block:
+
+- `package.json` — JS / TS framework detection (Next.js App vs Pages,
+  Remix, SvelteKit, Astro, Expo, …)
+- `composer.json` — PHP framework detection (Laravel, Symfony,
+  framework-less)
+- `pyproject.toml` / `requirements.txt` — Python framework detection
+- `CLAUDE.md` / `AGENTS.md` — project-declared stack hints
+- `.cursorrules` — project-declared stack hints
+- `tsconfig.json` — TS path-alias / module-resolution hints
+
+The skill MUST NOT read source files, `.env*`, secrets, or user
+data. Detection lands as a single `assumes: stack=<framework>@<version>`
+line; the medium-band halt is the user's chance to flip it.
+
 ## Procedure
 
 ### 1. Read and analyze the prompt
@@ -220,11 +264,24 @@ For `low`, the question replaces the AC list:
   `data.reconstructed_ac` and `data.assumptions`.
 - Do NOT re-derive band thresholds in prose. They live in
   `confidence.py` and only there.
+- Do NOT read source files, `.env*`, secrets, or arbitrary user
+  files when stack-detecting in mini / max mode. The allowlist
+  above (`package.json`, `composer.json`, `pyproject.toml`,
+  `requirements.txt`, `CLAUDE.md`, `AGENTS.md`, `.cursorrules`,
+  `tsconfig.json`) is exhaustive.
+- Do NOT strip the `bypass_prefix` mid-prompt. The prefix is only
+  recognised at the prompt boundary; matches inside the body stay
+  literal.
+- Do NOT silently rewrite the prompt in `max` mode without
+  surfacing the assumption list on a medium-band halt. The diff
+  is the contract.
 
 ## See also
 
 - [`refine-ticket`](../refine-ticket/SKILL.md) — sibling for ticket-shaped input
+- [`prompt-optimizer`](../prompt-optimizer/SKILL.md) — engine-outbound sibling; same `prompt_optimization` setting controls its mode
 - [`work_engine.resolvers.prompt`](../../templates/scripts/work_engine/resolvers/prompt.py) — envelope builder
 - [`work_engine.scoring.confidence`](../../templates/scripts/work_engine/scoring/confidence.py) — rubric + band thresholds
 - [`ask-when-uncertain`](../../rules/ask-when-uncertain.md) — one-question-per-turn Iron Law
 - [`artifact-drafting-protocol`](../../rules/artifact-drafting-protocol.md) — this skill was drafted under it
+- AI Council session: `agents/council-responses/prompt-master-mini.json` (2026-05-17) — analysis behind the mini/max split and `/raw` bypass <!-- council-ref-allowed: ADR decision trace -->

package/.agent-src/skills/tailwind-engineer/SKILL.md

@@ -117,6 +117,20 @@ Risks:          <arbitrary values, !important, dark-mode gaps>
   break the design system; they accumulate silently.
 - `@apply` inside component CSS interacts with PurgeCSS — keep it
   in files Tailwind scans, not in vendor CSS.
+- **Anti-AI-slop: gradients.** Unless audit-pinned or brief-explicit,
+  avoid the default purple-to-blue / cyan-to-pink gradients on white —
+  they read as auto-generated. Reach for a single accent from the
+  token map, or a duotone built from configured tokens.
+- **Anti-AI-slop: typography.** Unless audit-pinned, avoid surfacing
+  the system stack (`font-sans` fallback to Arial / Helvetica / Inter
+  via system defaults) as the *visible* body face. If `tailwind.config`
+  pins a font family, use it; if not, treat the missing token as a
+  gap to flag, not a license to ship the OS default.
+- **Anti-AI-slop: layout.** Unless audit-pinned, the centered hero +
+  3-column features + CTA stack is the AI-template tell. Break the
+  grid intentionally (asymmetric column split, overlap, diagonal
+  flow) when the brief allows; cite the design brief's `aesthetic:`
+  line if `fe-design`'s aesthetic-direction section produced one.
 
 ## Do NOT
 

package/.agent-src/templates/agents/agent-project-settings.example.yml

@@ -39,7 +39,7 @@ schema_version: 1
 # CI guard: a release bump of `package.json` must update this value
 # in lockstep — see scripts/check_template_pin_drift.py (road-to-
 # portable-runtime-and-update-check P3.3).
-agent_config_version: "2.20.0"
+agent_config_version: "2.21.0"
 
 # --- Project identity ---
 project:
@@ -152,6 +152,58 @@ quality:
     # Known: eslint, prettier, tsc, biome.
     tools: [eslint, prettier, tsc]
 
+# --- Prompt optimization (engine-inbound + engine-outbound) ---
+#
+# Controls how aggressively the agent reshapes a free-form prompt
+# *before* the engine plans (`/work "<prompt>"`) and how the
+# outbound `/optimize-prompt` skill polishes a copy-and-paste
+# prompt for an external AI.
+#
+# Single-knob model — `default:` applies to both inbound and
+# outbound surfaces unless an override key is set:
+#
+#   off  — engine runs the user's literal prompt. No reconstruction,
+#          no AC inference, no template selection. Closest to "raw"
+#          behaviour; useful for power users who already write
+#          structured prompts.
+#   mini — light shaping. Inbound: `refine-prompt` runs in
+#          stack-aware mode (reads `package.json` / `CLAUDE.md` /
+#          `.cursorrules` only to infer the framework), asks at
+#          most 5 targeted clarifying questions, emits a structured
+#          6-block prompt. Outbound: BASIC mode of `prompt-optimizer`
+#          (4-D pass with safe defaults, no clarifying questions).
+#   max  — full restructure. Inbound: `refine-prompt` runs the
+#          5-dimension confidence rubric + assumption inference;
+#          medium-band halts surface the assumption list, low-band
+#          asks one clarifying question. Outbound: DETAIL mode of
+#          `prompt-optimizer` with template selection (RTF, CO-STAR,
+#          RISEN, CRISPE, …) per `docs/guidelines/prompt-templates.md`.
+#
+# Default `max` reshapes every prompt before the engine plans —
+# the AI Council (anthropic/openai, 2026-05-17) warned this risks
+# latency, token cost, and loss of author intent on the inbound
+# side. If that bites you, uncomment `inbound: mini` below. The
+# outbound side (explicit `/optimize-prompt`) is opt-in and stays
+# at `max` regardless.
+#
+# Bypass: any prompt starting with `bypass_prefix` skips both
+# inbound and outbound shaping verbatim. `/`-prefixed slash
+# commands (`/work`, `/commit`, …) and `#`-prefixed memory entries
+# auto-bypass.
+prompt_optimization:
+  default: max
+
+  # Optional per-surface overrides. Leave commented to inherit
+  # `default:`. Council-recommended split is `inbound: mini,
+  # outbound: max` — uncomment if you want fewer interceptions
+  # on `/work` but full polish on `/optimize-prompt`.
+  # inbound: mini
+  # outbound: max
+
+  # Prefix that skips all shaping. Must start with `/` to align
+  # with this project's slash-command convention. Default `/raw`.
+  bypass_prefix: "/raw"
+
 # --- Locked keys (override this file only; never locks .agent-settings.yml) ---
 #
 # List keys from this file whose values cannot be overridden by a

package/.claude-plugin/marketplace.json

@@ -6,7 +6,7 @@
   },
   "metadata": {
     "description": "Shared agent configuration \u2014 skills for AI coding tools (Claude Code, Augment, Cursor, Cline, Windsurf, Gemini CLI).",
-    "version": "2.20.1",
+    "version": "2.23.0",
     "keywords": [
       "agent-config",
       "skills",
@@ -99,6 +99,7 @@
         "./.claude/skills/competitive-positioning",
         "./.claude/skills/composer-packages",
         "./.claude/skills/compress",
+        "./.claude/skills/compress-memory",
         "./.claude/skills/content-funnel-design",
         "./.claude/skills/context",
         "./.claude/skills/context-authoring",