HalalifyMusic · denfry · Jul 3, 2026 · Jul 3, 2026 · Jul 3, 2026 · Jul 3, 2026
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -7,6 +7,54 @@ follow [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 ## [Unreleased]
 
 ### Fixed
+- The shell profile sourced the launcher from the cloned repo, so deleting or
+  moving the clone broke every new shell, and a re-install from a new path
+  silently kept the stale line. The installer now copies `fable.ps1`/`fable.zsh`
+  into `~/.claude/shell/` and sources the stable copy; a stale launcher line is
+  replaced, not kept.
+- Auto/effort-activated sessions lost the playbook forever after a compaction:
+  SessionStart re-injected only under `FABLE_MODE=1`, while the session marker
+  blocked the UserPromptSubmit path. The trigger hook now re-injects on
+  `source == "compact"` whenever the marker proves the mode was active, and
+  drops the marker on `source == "clear"` so the once-per-session paths re-arm.
+- `FABLE_PLAYBOOK.md` instructed the model to read the author's personal files
+  (`~/Downloads/Fable_Mindset_public.md`, `~/compare_models.py`,
+  `~/fable_dataset_delta.py`, `reference/llm-bias-awareness.md`) that no user
+  has — every fable session risked a failing Read. They are now provenance
+  notes, not paths; a content-guard test keeps them out.
+- The `/fable` skill's fallback ("read the playbook from this skill's
+  repository") was unreachable after install — the installed skill is a lone
+  SKILL.md. The fallback now points at `fable doctor` / re-running the
+  installer.
+- The installer silently `rmtree`'d a user's own `~/.claude/skills/<name>`
+  when it collided with a bundled skill name; it now preserves such
+  directories under `~/.claude/backups/skills/` and marks its own copies with
+  a `.fable-mode-bundled` file. The uninstaller removes only marked
+  directories.
+- `test-after-edit.py` ran the suite even when the Edit tool itself had
+  failed (reporting a stale result), and `fable.ps1 doctor` hardcoded
+  `python`, breaking on machines where only `py`/`python3` exists. Both
+  fixed; the doctor subcommand now falls back across `python`/`py`/`python3`.
+- Session/debounce marker files accumulated in `%TEMP%` forever on Windows;
+  both hooks now garbage-collect markers older than a week.
+- "Invalid JSON provided for --settings" on Windows PowerShell 5.1 (#2): PS 5.1
+  strips embedded quotes when building the native command line, so the inline
+  `--settings '{"ultracode": true}'` reached claude as `{ultracode: true}`. The
+  launchers now pass a settings *file* (`~/.claude/ultracode.settings.json`,
+  shipped and installed), which survives quoting on every shell and PS version.
+- The playbook injector never actually fired: the effort path needs `effort` in
+  the hook payload or `CLAUDE_EFFORT` in the hook environment, which Claude Code
+  ≤ 2.1.198 did not provide, and the trigger phrases were undiscoverable. The
+  launcher now declares the mode via `FABLE_MODE=1` and `fable-trigger.py`
+  became dual-event — SessionStart injection (version-independent) plus the old
+  phrase/effort paths and a `FABLE_MODE` fallback on UserPromptSubmit.
+- `merge_settings.py` overwrote `settings.json.bak` on every install run, so a
+  second install destroyed the pristine pre-install backup (contradicting the
+  "safe to re-run" promise). It now writes that backup only when absent, keeping
+  the original restore point intact.
+- Closed leaked file handles (`json.load(open(...))` / `json.dump(..., open(...))`)
+  in `merge_settings.py`, `uninstall.py`, and `test-after-edit.py` — they now use
+  `with` blocks, which also avoids file-lock surprises on Windows.
 - `fable-trigger.py` read the playbook from a hardcoded `/Users/ak/...` path, so
   on-demand injection silently failed for everyone but the original author. It now
   resolves `~/.claude/FABLE_PLAYBOOK.md`.
@@ -16,6 +64,43 @@ follow [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
   `.lockb` skip entry.
 
 ### Added
+- Double-injection guard: before the once-per-session auto/effort injection,
+  the trigger hook scans the session transcript for a "Fable mode active"
+  line — a session already activated via the `/fable` skill doesn't get a
+  second copy of the playbook.
+- On Windows the launcher is now installed into the `$PROFILE` of **both**
+  PowerShell 7 and Windows PowerShell 5.1 when both are present.
+- `fable doctor` now also verifies the launcher line in the shell profiles
+  (a line pointing at a missing file is a hard failure), accepts hooks
+  registered in `settings.local.json`, probes `claude --help` for the
+  `--effort` / `--append-system-prompt-file` flags the launcher relies on,
+  and checks the installed `~/.claude/shell/` launcher copies.
+- The auto-activation heuristic learned more everyday task verbs
+  (update/upgrade/remove/delete/deploy/rename/write; удалить/обновить/
+  написать/запустить/перенести/убрать).
+- Auto-activation: `fable-trigger.py` scores prompt complexity (ru+en signals:
+  task verbs, code fences, file paths, multi-step markers, length) and loads
+  the playbook by itself for task-shaped prompts, once per session, in any
+  session — no launcher or phrase needed. Opt out with `FABLE_AUTO=0`.
+- `fable --ultra` (alias `-u`): launches with ultracode auto-orchestration; new
+  "Orchestration" section in the playbook (fan-out, adversarial verification
+  via `grounding-verifier`, plan-gating, calibration).
+- `fable doctor`: one-command diagnosis of the install/activation chain —
+  files, registered hooks, interpreter paths, Claude Code version, a live-fire
+  injection test, and transcript evidence of past activations.
+- The `/fable` skill triggers proactively at the start of non-trivial tasks.
+- `fable-code.md`: an original Claude Code-native Fable behavior layer; the
+  launcher appends it instead of the 1,600-line consumer prompt (which stays
+  bundled for reference).
+- `/fable` skill for explicit mid-session activation.
+- The launcher pins `--model claude-opus-4-8`, making the README's "runs on
+  Opus 4.8" promise true regardless of the user's default model.
+- Round-trip uninstall test (`tests/test_uninstall.py`).
+- `test-after-edit.py` gained a `FABLE_TEST_HOOK_ALLOW` allowlist (os.pathsep-
+  separated trusted root prefixes) so the auto-run test hook can be confined to
+  repositories you trust, plus a per-project `.fable-test` file to pin the exact
+  command it runs (e.g. a fast, scoped command instead of the whole suite in a
+  monorepo). Documented in `SECURITY.md`; covered by new tests.
 - **One-command, cross-platform installer** (`install.py`) for Windows, macOS, and
   Linux. `install.sh` / `install.ps1` are thin wrappers that exec it.
 - PowerShell launcher (`shell/fable.ps1`) alongside the zsh one.
@@ -28,6 +113,11 @@ follow [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
   `SECURITY.md`, `CONTRIBUTING.md`, and this changelog.
 
 ### Changed
+- The `fable` launcher now defaults to `--effort xhigh` instead of
+  `--settings '{"ultracode": true}'`. xhigh still trips the playbook trigger and
+  drives heavy reasoning, without ultracode's token-hungry multi-agent
+  auto-orchestration; ultracode is now an explicit opt-in (documented in the
+  launcher and README).
 - `merge_settings.py` writes the absolute interpreter (`sys.executable`) and
   absolute hook paths into `settings.json`, so the hooks fire without `$HOME` or
   `python3` resolution at hook-run time.

diff --git a/FABLE_PLAYBOOK.md b/FABLE_PLAYBOOK.md
@@ -1,14 +1,15 @@
 # Execution Playbook — Fable-5 patterns for Opus 4.8 (measured companion)
 
-This is the **measured** companion to *The Fable Mindset* (`~/Downloads/Fable_Mindset_public.md`).
-That document is the full disposition manual; read it first. This file does one
-thing it doesn't: it puts **your** fable-5 turns next to **your** opus-4-8 turns,
-head to head, so every "adopt this / refuse that" call is backed by a number
-from your own history — and it is honest about where the gap was smaller than it
-first looked.
+This is the **measured** companion to *The Fable Mindset*, the community
+disposition manual (not bundled here — nothing below depends on having it).
+This file does one thing that manual doesn't: it puts real fable-5 turns next
+to real opus-4-8 turns, head to head, so every "adopt this / refuse that" call
+is backed by a number — and it is honest about where the gap was smaller than
+it first looked.
 
-Measured: 1,307 fable-5 turns vs 10,470 opus-4-8 turns across 139 / 243 sessions.
-Re-run anytime with `~/compare_models.py`.
+Measured: 1,307 fable-5 turns vs 10,470 opus-4-8 turns across 139 / 243
+sessions. The measurement tooling is not part of this bundle; the numbers are
+the deliverable.
 
 ---
 
@@ -60,8 +61,9 @@ work (mostly write-heavy greenfield builds). What it confirms and what it compli
   chain-of-thought on every event (100%); your Opus logs only store visible
   thinking blocks. Don't read a reasoning delta off these two columns.
 
-Ruler: `~/fable_dataset_delta.py`; extracted profile: `~/fable5_dataset_profile.json`.
-The 22 MB raw dataset was deleted after the metrics were extracted.
+The ruler script and extracted profile are part of the measurement tooling,
+not this bundle. The 22 MB raw dataset was deleted after the metrics were
+extracted.
 
 ---
 
@@ -203,8 +205,8 @@ makes the work correct; this layer makes the output read like Fable wrote it.
 - **Own mistakes without grovelling.** When wrong, acknowledge it plainly, fix it,
   stay on the problem — no self-abasement, no excessive apology, no caving on a
   correct position just because you were challenged. Maintain self-respect; the
-  goal is steady honest helpfulness. (Pairs with `reference/llm-bias-awareness.md`:
-  a challenge is not an automatic signal to surrender.)
+  goal is steady honest helpfulness. A challenge is not an automatic signal to
+  surrender.
 - **Epistemic honesty / no confabulation.** Partial recognition from training is
   *not* current knowledge. For any library, version, product, or fact that may have
   moved, verify (Context7 / web / the file itself) rather than answering from
@@ -232,6 +234,24 @@ makes the work correct; this layer makes the output read like Fable wrote it.
   view of that file is stale — re-read the region before editing it again. (Extends
   ADOPT 3 past the first edit; Fable states this outright for `str_replace`.)
 
+### The Claude Code layer (first-party, 2026-07)
+
+Observed first-party Fable 5 behavior in Claude Code adds three corrections to
+the voice layer above — adopt them with the same weight:
+
+- **The final message is the deliverable.** Mid-turn text may never be shown;
+  answers, findings, and conclusions must all appear in the turn's last
+  message, outcome first. A perfect investigation with a buried conclusion
+  reads as no conclusion.
+- **Readable beats concise.** Shorten by selecting what matters, not by
+  compressing prose into fragments, arrow chains ("A → B → fails"), or
+  shorthand the reader must reverse-engineer. Complete sentences; explain in
+  place. If the reader must reread or ask a follow-up, the brevity saved
+  nothing.
+- **Assessment mode.** When the user is describing a problem or thinking out
+  loud rather than requesting a change, the deliverable is the assessment:
+  report findings and stop; don't apply the fix until asked.
+
 ---
 
 ## Grounding — prove it before you call it done
@@ -305,18 +325,40 @@ levers the harness actually enforces:
   that runs the project's test command, with `hooksEnabled: true`. This is the
   fix for Fix 1; it fires every time, not most of the time.
 - **Voice/formatting** → not hook-enforceable; it is pure disposition. The closest
-  lever is the bundled `stop-slop` skill — invoke it for any substantial prose
-  (docs, PRs, commit messages, user-facing summaries). For everything else the
-  voice-layer section above is the standing rule.
+  lever is a `stop-slop` skill, if one is installed — invoke it for any substantial
+  prose (docs, PRs, commit messages, user-facing summaries). For everything else
+  the voice-layer section above is the standing rule.
 - **Grounding / verification independence** → the protocol ships as the `/ground`
   skill (`~/.claude/skills/ground/`, invoke by name for the full self-terminating
   ledger loop) and the cold `grounding-verifier` agent (`~/.claude/agents/`, spawn for
   an independent read-only check that never sees your reasoning). Fable mode loads the
   Grounding section above automatically via this playbook, so the standing discipline
   is always on; the skill and agent are the heavier, explicit forms. For code changes
-  the bundled `/code-review` + `/verify` skills run alongside the verifier. Never
-  self-approve in the same active context.
+  Claude Code's built-in `/code-review` + `/verify` skills run alongside the verifier.
+  Never self-approve in the same active context.
 - **Placement** → these rules belong in a `CLAUDE.md` (loads every session), not
   auto-memory (relevance-gated, may not surface on a given turn). Point sessions
   at this file and *The Fable Mindset* deliberately; wire the hook and effort
   level as the hard guarantees.
+
+---
+
+## Orchestration — scale the harness to the task
+
+Fable's discipline is single-context by default; escalate deliberately:
+
+- **Fan out when the task decomposes.** Independent units (N files to migrate,
+  M subsystems to map, review dimensions) → parallel subagents, each with a
+  scoped brief; synthesize in the main context. Never serialize what has no
+  data dependency (Fix 3, applied to agents).
+- **Verify adversarially.** Fan-in results are claims, not facts: route
+  non-trivial ones through the evidence ledger, and spawn the cold
+  `grounding-verifier` on the merged result — a subagent never self-approves
+  its own output (Grounding section above).
+- **Plan-gate before long autonomy.** Multi-phase runs get a phased plan and a
+  live task list first; return to the plan at each phase boundary.
+- **Calibrate ruthlessly.** One context that fits the whole task beats any
+  orchestration. No multi-agent machinery for typo-class work.
+- **Mechanical lever:** `fable --ultra` launches with `ultracode` — the harness
+  auto-runs multi-agent workflows for substantive tasks. Heavy on tokens; the
+  calibration rule above is the counterweight.
diff --git a/README.md b/README.md
@@ -39,7 +39,7 @@ python install.py        # Windows  (use python3 on macOS / Linux)
 Then reload your shell and launch:
 
 ```sh
-fable        # Opus 4.8 + Fable prompt + ultracode
+fable        # Opus 4.8 + Fable prompt + xhigh effort
 ```
 
 Prefer a native one-liner? `./install.sh` (macOS / Linux) and `.\install.ps1` (Windows) just locate Python and run `install.py` for you.
@@ -48,24 +48,26 @@ Prefer a native one-liner? `./install.sh` (macOS / Linux) and `.\install.ps1` (W
 
 > **Requirements:** Python on PATH (`python --version`) for the hooks, and Claude Code installed for the `fable` launcher. On Windows, if `.\install.ps1` is blocked by execution policy, run `python install.py` directly (no policy needed).
 
-The installer copies everything into `~/.claude`, adds the `fable` launcher (to your shell rc on Unix, to your PowerShell `$PROFILE` on Windows), and merges your settings (with a backup) — writing the absolute interpreter and hook paths so the hooks fire on every platform. Idempotent: safe to re-run. Needs Python ≥ 3.9 (the hooks are stdlib-only — no pip installs). No model switch, no API key — it runs on the Opus 4.8 you already have.
+The installer copies everything into `~/.claude` — including the launcher itself (`~/.claude/shell/`), so you can delete the clone afterwards — adds the `fable` launcher (to your shell rc on Unix, to the `$PROFILE` of both PowerShell 7 and Windows PowerShell on Windows), and merges your settings (with a backup) — writing the absolute interpreter and hook paths so the hooks fire on every platform. Idempotent: safe to re-run, and a same-named skill you wrote yourself is preserved under `~/.claude/backups/skills/` instead of being overwritten. Needs Python ≥ 3.9 (the hooks are stdlib-only — no pip installs). No model switch, no API key — it runs on the Opus 4.8 you already have.
 
 ## Uninstall
 
 ```sh
 python uninstall.py        # Windows  (python3 on macOS / Linux; or ./uninstall.sh, .\uninstall.ps1)
 ```
 
-Removes the bundled files from `~/.claude`, strips the `fable` launcher line, and drops the two Fable hooks from `settings.json` (writing a fresh backup). It's surgical — your own skills, unrelated hooks, and `alwaysThinkingEnabled` are left untouched.
+Removes the bundled files from `~/.claude`, strips the `fable` launcher line, and drops the two Fable hooks from `settings.json` (writing a fresh backup). It's surgical — only skill directories carrying the installer's `.fable-mode-bundled` marker are removed; your own skills, unrelated hooks, and `alwaysThinkingEnabled` are left untouched.
 
 ## What's in the bundle
 
 - **`FABLE_PLAYBOOK.md`** — the core. Fable-5 vs Opus-4-8 tool traces turned into rules: reasoning density (70% vs 47%), verify-after-edit, parallelism — plus a voice layer and an evidence-ledger grounding protocol. Original work, not the leaked prompt.
 - **`fable-system.md`** — the leaked Fable 5 system prompt (Anthropic's; see note below).
-- **Hooks** — `fable-trigger.py` injects the playbook at `xhigh`/`max`/`ultracode`; `test-after-edit.py` runs your project's tests after each edit and reports the result back — the one habit no model keeps on willpower.
+- **`fable-code.md`** — an original Claude Code-native distillation of Fable's terminal behavior (final-message contract, readable-over-concise, tool discipline, autonomy rules). This is what the launcher actually appends; the consumer prompt above stays bundled for reference.
+- **`/fable` skill** — explicit activation: reads the playbook + behavior layer and adopts both mid-session, no launcher required.
+- **Hooks** — `fable-trigger.py` injects the playbook when the launcher declares the mode, at `xhigh`/`max`/`ultracode` effort, on a trigger phrase, or **by itself when the prompt looks like a real task** (ru+en heuristic; opt out with `FABLE_AUTO=0`); `test-after-edit.py` runs your project's tests after each edit and reports the result back — the one habit no model keeps on willpower.
 - **`/ground` skill + `grounding-verifier` agent** — a self-terminating grounding loop and a cold verifier that assumes every claim is wrong until the live code proves it.
 - **Skills** — `claude-design-patterns` (web-UI engineering), `webapp-testing`, `mcp-builder`, `skill-creator`, `explore-data`.
-- **`fable` launcher** — Opus 4.8 + the prompt + `ultracode` effort (`fable.zsh` for Unix shells, `fable.ps1` for PowerShell).
+- **`fable` launcher** — pins `--model claude-opus-4-8`, appends `fable-code.md`, sets `xhigh` effort, and declares the mode via `FABLE_MODE=1` so the playbook injects at session start on every Claude Code version (`fable.zsh` for Unix shells, `fable.ps1` for PowerShell). `fable --ultra` adds ultracode multi-agent orchestration; `fable doctor` verifies the whole install/activation chain in one command.
 
 ## The honest ceiling