v1.7.0 — the closed loop: it learns your model's bad habits

• Closed-loop habit correction: the Stop hook detects discipline violations from session transcripts (re-reads after edits, repeat reads, full-file rewrites, unverified edits); the next session gets ONE targeted corrective for the worst recurring habit. Local-only, one line of context.
• Marathon tier × 3 at n=2 (Sonnet): 12/12 pass both conditions; fable-mode −35% report length, −14% turns, −8% output tokens. Haiku fable-condition record: 9/9 all-time.
• Grading self-test now covers every task (20 checks).

v1.6.1 — breakpoint clearing, provider-neutral export

• Breakpoint clearing: the state loop suggests /clear at task boundaries (done + ctx >40%), not on a timer — cached history is cheap; rebuilding context isn't. New ⟳ clear-worthy statusline chip.
• Provider-neutral wording: the AGENTS.md export now degrades gracefully on harnesses without subagents/compaction — Claude, GPT, Gemini, or local models.
• Regression evidence: Haiku 7/7, first perfect sweep including the 12-file marathon task.

v1.6.0 — the state loop, marathon evals, prompt-variant testing

The state loop

/fable-mode:fable-state on — Fable maintains a .fable-state.md project memory file at the end of every substantive turn. /clear whenever the context meter runs hot; the next session auto-resumes from the file. Continuation context in ~2k tokens instead of re-billing your entire chat history every turn. Off by default, statusline shows ✎ state while active.

Also in this release

• Marathon eval tier: 12-file service fixture with a cross-file root cause — measures reading discipline where it actually compounds.
• --skill-file flag: A/B alternate SKILL.md variants; prompt changes are now measured, not guessed.
• fable-mode-skill.zip attached below — one-step claude.ai skill upload.
• /fable-stats estimated-savings line, README visuals, CHANGELOG, CONTRIBUTING, issue templates, Windows notes.

v1.5.1 — eight more craftsman rules

Skill additions, all woven into the existing loop:

• One line? One line. And between two equal-size options, take the one that's correct on edge cases — less code never means a flimsier algorithm.
• Deletion over addition, boring over clever, and prefer editing an existing file over creating a new one.
• User insists on the full version? Build it — no re-arguing.
• YAGNI applies to tests too: one minimal check, no frameworks or per-function suites unless asked.
• 'skipped X — add when Y' one-line reporting convention for deliberate omissions.
• Requested explanation is never debt — brevity rules bind only unrequested prose.

AGENTS.md regenerated to match.

v1.5.0 — usage-limit meters, setup guide, the Fable persona

• Usage-limit meters in the statusline: native 5-hour and 7-day plan quota bars (Pro/Max) with reset times — yellow at 50%, red at 80%. Context meter now prefers Claude Code's official context_window fields with a transcript fallback for older CLIs.
• INSTALL.md: complete setup guide — plugin, statusline, level pinning, stats, claude.ai upload, AGENTS.md export, updating, uninstalling, troubleshooting.
• The plugin is now 'Fable' (displayName): a persona — active every response, released with fable off.
• Sharper craftsman mechanics in the skill: rung zero ('does this need to exist at all?'), fable: shortcut comments naming the ceiling and upgrade path, ship-the-minimal-version-and-offer-the-full-one, and an anti-drift persistence clause.

v1.4.0 — statusline HUD + /fable-level

• Statusline HUD: opt-in terminal status bar showing fable level, model, branch, live context meter with tokens remaining (yellow 60% / red 85%), session output tokens, cost, duration, lines changed. Install: scripts/install-statusline.sh (refuses to clobber an existing statusline without --force). ~50ms render, local files only.
• /fable-level command: pin lite|full|deep or return to auto from chat; persists across sessions.
• SessionStart hook now keeps the installed statusline copy in sync with plugin updates.

v1.3.0 — auto-routed levels, local stats, portable export, reproducible evals

What's new

• Intensity levels with auto mode — fable-mode silently routes each task to lite / full / deep; no classification needed from the user. Pin a level in chat (fable level: deep) or persistently via ~/.claude/fable-mode/level. Misrouted tasks escalate mid-task.
• Local token stats — a Stop hook logs per-session usage to the plugin data dir (local only, no telemetry). Ask /fable-mode:fable-stats for totals and trends.
• AGENTS.md export — the same operating style, generated from the skill, for Cursor / Codex / Zed / any AGENTS.md-aware tool.
• Reproducible eval harness — evals/ ships six graded A/B tasks (root-cause instinct, triage, reuse, YAGNI, two green-test-suite traps) plus a stdlib-only runner with real token accounting, and a 14-check grading self-test.
• CI — manifests, hook scripts, skill frontmatter, AGENTS.md sync, and the grading self-test validate on every push.
• Skill additions: prompt-cache hygiene, long-session state notes, numeric uncertainty calibration.