bossbuild 0.97.0

package/registry/CHANGELOG.md

@@ -0,0 +1,2737 @@
+# BOSS Changelog
+
+Each entry = a BOSS version. `/boss-sync` reads this to tell a project what's new since its pin.
+
+## 0.97.0 — 2026-06-23
+
+- **Rebrand: "BlueprintOS" → BOSS (Build Out Solid Stuff).** The name "BlueprintOS" collided with
+  [gj1342/blueprint-os](https://github.com/gj1342/blueprint-os) (an "OS for AI agents" using the same
+  skills/specs-as-markdown vocabulary), so the long form is retired and the earned **BOSS** wordmark
+  re-founded: **B.O.S.S. = Build Out Solid Stuff**, slogan **"Make it real."** (the anti-pseudo-app thesis
+  in three words). Dropped the "Operating System" framing entirely — BOSS is an incubator with a conscience,
+  not an OS. **Renamed across the shipped surface:** npm package `blueprintos` → **`bossbuild`**, repo URLs →
+  `github.com/ajeshh/bossbuild`, README / PRINCIPLES / docs headers, scaffold templates ("Scaffolded by
+  BOSS"), and CLI version/working-rules strings. **The `boss` command, the voice, the `mentor-*` roster, and
+  PRINCIPLES are unchanged** (package ≠ command — the bin name is always ours). Decision recorded in DEC-002;
+  full identity in `docs/design/BRAND.md`. Easter-egg full forms: *Builds, Or Stays Silent* · *Bosses Only
+  Self-Sabotage*. **Next:** GitHub repo rename, npm publish as `bossbuild`, Homebrew tap, domain `boss.build`.
+
+## 0.96.0 — 2026-06-21
+
+- **First `/humane-refresh` sweep, pass 2 → the cohort & frontier patterns (RVW-059–064).** The second
+  `/deep-research` pass (111 agents, 24/25 claims 3-vote verified) covered the families the AI-chatbot lens
+  and the classic-web checklist both miss. Folded into
+  [`ai-ux-patterns.md`](../library/practices/ai-ux-patterns.md) as a new **"cohort & frontier patterns"**
+  block (conditional — surfaces when the product touches that surface, not on every Quickstart):
+  - **Accessibility / exclusion-by-design (RVW-059)** — an inaccessible flow (unlabeled element, inaccessible
+    CAPTCHA, screen-reader-invisible pre-ticked box) is *"effectively deceptive"* to anyone who can't perceive
+    or escape it, **even unintentionally** — the sharpest case of "effect, not intent." WCAG is the floor, not
+    the ceiling; test deceptiveness *under assistive tech*. (CHI/CSCW '25.)
+  - **Minors & vulnerable-by-design (RVW-061)** — price in real currency + disclose odds (loot-box/gacha
+    currency-laundering; FTC's $20M Genshin action), ship addictive-design OFF by default for minors
+    (streaks/autoplay/push — EU DSA), age assurance not age-gate theater. (UK Children's Code, statutory.)
+  - **Agentic dark patterns (RVW-060)** — *the keystone for BOSS, which builds agents.* Two directions: your
+    agent as **perpetrator** (commit/purchase without consent, over-broad permissions → scope + confirm, §4's
+    risk gate) and your agent as **victim** (manipulated by UI dark patterns **70%+ vs 31% for humans, worse
+    as models scale** — Stanford DECEPTICON / OWASP Agentic 2026). The victim half also hardens
+    [`agent-security.md`](../library/practices/agent-security.md): UI dark patterns are an injection surface,
+    and **recognition ≠ protection** (in-context prompting, guardrail models, *and human oversight* were each
+    shown insufficient — don't oversell a fix).
+  - **Algorithmic management (RVW-063)** — when a product scores/ranks/pays people, opaque unpredictable
+    scoring + "algorithmic gamblification" is the worker-facing dark pattern → transparent predictable
+    formula. (HRW *The Gig Trap* 2025.)
+  - **Junk fees (RVW-062)** sharpens the v0.95 drip-pricing line — the all-in total must be the *most
+    prominent* price (FTC Junk Fees Rule). Regulatory-teeth pointer extended (Genshin, EU DSA minors, UK
+    Children's Code, ADA/EAA) **with an honest caveat that the FTC click-to-cancel rule was vacated (8th Cir.
+    2025).**
+  - **Sludge (RVW-064) held NOT-YET** — Family 1 returned zero verified claims (its *mechanics* already
+    landed under Obstruction in v0.95); needs a dedicated pass, alongside the click-to-cancel status. Watchlist
+    edges updated.
+  - *Founder-facing, project-neutral; pulled via `/boss-sync`.*
+
+## 0.95.0 — 2026-06-21
+
+- **`/humane-refresh` — the humane lens stops being a snapshot (IDEA-042).** BOSS's dark-pattern catalog was
+  adopted once (the CDT 37, v0.82) and frozen. Dark patterns are an arms race — new research and regulation
+  name them, new models open new emergent surfaces, new tools (generative UI, agents-on-your-behalf) create
+  surfaces nobody's named. So the catalog now has a **standing freshness discipline** instead of a one-time
+  read. The staleness-twin of model recalibration ([IDEA-014](../docs/ideas/IDEA-014-model-recalibration-discipline.md));
+  they share trigger events.
+  - **New skill [`.claude/skills/humane-refresh`](../.claude/skills/humane-refresh/SKILL.md)** (a
+    BOSS-curating-BOSS meta-skill, sits with `/vet`/`/boss-learn`/`/boss-sync`). It *orchestrates skills that
+    already exist* — `/deep-research` **finds** (scoped since last run) → diff against the live catalog →
+    `/vet` **judges** (default NO) → `/boss-learn` **routes**. Three triggers: on-demand, quarterly via
+    `/schedule`, or `--event` for a new model / tool / regulation. Honest about what it can't auto-detect
+    (the event trigger is you running it, not performed magic). **Re-runs research, never freezes a list**
+    (the RVW-001 anti-rot guard made into a skill).
+  - **New [watchlist](../docs/research/watchlists/humane-lens.md)** — the focused half of `SOURCES.md`: the
+    regulators + canonical taxonomies it lacked (FTC, EDPB, CPPA, DSA, Brignull, Mathur, CHI/SOUPS/FAccT),
+    the standing query, and a `last_refresh` marker so each sweep scopes to "since last time."
+- **First sweep, pass 1 → the classic-web patterns the AI-chatbot lens omits (RVW-056/057).** Two
+  `/deep-research` passes, adversarially verified (3-vote). Folded into
+  [`ai-ux-patterns.md`](../library/practices/ai-ux-patterns.md):
+  - **Four named families an AI product inherits the moment it has an account, a paywall, or a checkout** —
+    **Obstruction** (roach-motel / un-deletable accounts, hard-to-cancel), **Sneaking** (sneak-into-basket,
+    hidden costs, drip pricing), **Manufactured urgency & scarcity** (fake timers/low-stock), **Interface
+    interference** (visual misdirection, trick questions, bad defaults) — each with its humane alternative.
+    Canon pinned (Brignull / Mathur / Gray), not enumerated (anti-freeze).
+  - **"Effect, not intent"** — a dark pattern needs no malice (CCPA judges by effect); you can ship one by
+    accident, which is exactly where the conscience earns its keep.
+  - **"Symmetry in choice"** as the concrete, testable bar under the existing asymmetry principle (CCPA
+    § 7004: the good door can't take more clicks than the bad one).
+  - **Regulatory teeth as a reference pointer** (CCPA, EU AI Act Art. 5, EDPB, FTC) — "is this regulated?",
+    explicitly *not* legal advice, *not* a gate.
+  - Pass-1 *under-sourced* claims (token-cost surprise, AI-washing, deepfake social proof, FTC
+    click-to-cancel specifics) held at **NOT-YET** (RVW-058) — named in blogs, not yet verifiable; the
+    watchlist re-sweeps them. Pass 2 (sludge, accessibility, minors, agentic-AI, financial, attention/labor)
+    folds in next.
+  - *Founder-facing, project-neutral; pulled via `/boss-sync`.*
+
+## 0.94.0 — 2026-06-21
+
+- **The generative half of the humane lens — first additions (composted from the humane-tech corpus).**
+  BOSS's humane lens has been *defensive* (harm-taxonomy, dark-patterns, conscience-not-censor = *don't
+  harm*). This batch begins the *generative* half — *cultivate flourishing* — drawn from a founder
+  humane-tech corpus (the Humane Product Canvas lineage, permaculture, the "Celebration of Done," play,
+  kinship). All judgment/voice/default touches, no new gates.
+  - **New [`library/practices/celebration-of-done.md`](../library/practices/celebration-of-done.md) +
+    wired into `/close`.** BOSS *records* done everywhere and *marks* it nowhere. Done is a threshold, not
+    perfection: a pause that registers what was crossed (against AI-speed build-amnesia), re-anchors on the
+    *why* and the *who*, and turns into curiosity about whether it resonates *now* — the bridge back to the
+    real user. The hard rule: celebrate **without performed warmth** ("🎉 great job!" is voice-mode bleed) —
+    genuine, specific, proportional, no streaks. It's a conscience moment with the polarity flipped.
+  - **`/prototype` gains the *magic circle*.** Named BOSS's one licensed *play* space — bounded,
+    safe-to-fail, nothing precious, the conscience waits outside. The generative register BOSS lacked.
+  - **`conscience-voicing` — not-knowing as a doorway.** Folded into the never-imply-an-intelligence-gap
+    rule: a founder's gap is an invitation ("haven't learned to appreciate it *yet*"), never a deficit.
+  - **`/canvas` Metrics cell — regenerative metrics, weighted as seriously as commercial:** what people
+    *learn* / gain in well-being, what makes the venture *resilient* — growth that renews, not just extracts.
+  - **`ai-ux-patterns` — "Humane defaults" (the build-time inverse of the dark-pattern checklist).**
+    Friction-as-ethics, done right: ship the humane choice as the default (remove friction from the good
+    path), **never** sabotage the other path (that's a dark pattern aimed at a goal you like — means matter,
+    sovereignty is non-negotiable, keep every door open). When the founder crosses anyway: name the cost
+    once, then **record the crossing** as a `DEC-NNN` — accountability, not a gate, so "why did we do that,
+    and when?" stays traceable. The antidote to humanity eroding *invisibly* through a thousand small
+    decisions.
+  - *Founder-facing, project-neutral; pulled via `/boss-sync`. The defensive + generative halves now both
+    ship.*
+
+## 0.93.0 — 2026-06-21
+
+- **Distribution as a voicing leg — `/ship` now asks "who finds it?" (IDEA-041, the FEAT-024 sequel).**
+  Closes a real asymmetry: `PRINCIPLES.md` names *path to distribution* as a co-equal leg of a real-value
+  app, but it was the only leg the conscience never voiced — *"will anyone pay?"* fires in the flow (moment
+  #1 + `mentor-business`), *"will anyone ever find it?"* didn't. **Decided the leg-vs-mentor question with
+  Ajesh:** not an auto-firing hook (its predicate needs the reachability detection deliberately parked in
+  FEAT-024 slice 3 — building it would be the over-fire nag that repels the anti-growth-hacking cohort
+  IDEA-019 protects), and not left as mentor-only (the gap stays). Instead a **third answer**: a single
+  **demand voicing wired into `/ship`** at the live moment — *"it's reachable now; who's the first real user,
+  and how do they hit this?"* Leg-like (fires in the flow, at the right moment, names the cost) without a
+  hook (rides the founder's deliberate `/ship` invocation → zero over-fire). Governed by IDEA-019's
+  **situation-not-person** rule (about the work's path to a user, never a judgment of the founder), kept to
+  the *demand* question (not a Product-Hunt checklist — that's the growth-hacking nag), once + suggestive +
+  points at `mentor-gtm` for depth, never a gate. Cohort-aware (lightest touch for `indie-hacker` /
+  anti-growth-hacking; skip entirely if they've a first user or a deliberate no-distribution stance).
+  `ship-it-live` practice gains a short "reachable → discoverable" note. `/tmp`-verified (skill scaffolds
+  clean, 0 placeholders). **The full distribution *hook* stays deferred** — earned only if a reliable
+  reachable-but-undiscovered predicate ever exists (shares FEAT-024 slice 3's parked substrate question).
+
+## 0.92.0 — 2026-06-21
+
+- **`ship-it-live` practice + `/ship` skill — the CD half of building (FEAT-024 slices 1+2).** `git-workflow`
+  shipped the CI half ("is `main` green"); this ships the **CD** half: *is this where a real user can hit it,
+  or just you?* An app that only runs on `localhost` is a pseudo app — the validation conscience has no teeth
+  if the artifact was never put where someone could prove it. Shaped past "starter spec" by a `/deep-research`
+  pass (21 sources, 25 claims adversarially verified 3-vote, 22 confirmed / 3 killed).
+  - **UP — new [`library/practices/ship-it-live.md`](../library/practices/ship-it-live.md)** (stack-neutral,
+    Principle #4 — no baked-in deploy target): **"localhost is not shipped"**; **deploy early/cheap/reversible**
+    (the *"reliability is premature at MVP"* counter was killed 0-3 — the headline stands un-hedged);
+    **secrets & authz at the boundary = the leg with teeth** — the signature vibe-coded failure is a
+    client-bundled DB key + RLS the AI never configured (**CVE-2025-48757 / Lovable** — 170+ apps, ~10.3%
+    leaking PII + keys; **MoltBook** — 1.5M credentials, founder wrote no code) [EVIDENCE]; **rollback ≠
+    reversible** (instant rollback restores the *app*, not the *database* — Vercel documents this against its
+    own feature → expand-migrate-contract, Fowler); a **deploy honesty anchor** (DORA 2024: AI ↔ *worse*
+    stability + throughput — the METR twin; small-batches-offset-it was killed 1-2, so it's stated not
+    oversold); **preview-per-branch demoted** from "the review primitive" to a JIT-gated judgment (its
+    loop-tightening efficacy is vendor-positioning only — the one verified counter-argument; over-ceremony for
+    a 2-person team).
+  - **DOWN — new `/ship` skill (L1/MVP)** — the deterministic "put it where a real user can hit it" runner:
+    detect the stack → **deploy-time pre-flight** (no client-bundled secrets; server-side authz/RLS actually
+    on) → cheapest reversible host → hand back the **live URL** + the rollback path → capture the recipe as a
+    `PRAC-NNN`. The pre-flight is a **check, not a gate** (conscience-not-censor) and **points at** `/red-team`'s
+    pre-ship pass + `agent-security` rather than restating them (the secret-scan surface `secrets-guard`
+    deliberately doesn't cover). Cohort-aware (the pre-flight is non-negotiable for `first-product` /
+    `non-tech-founder` — they can't spot a leaked key). Plus a tight **"Shipping (localhost is not shipped)"**
+    DOWN section in the L1/MVP working rules. Skill-layer (predicate/runner split — the model parses the stack
+    + shells out, so `src/` stays zero-dep, Principle #4 / working-rule #4).
+  - **Slice 3 (the `reachable?` conscience moment) deliberately deferred** — voicing-first; a hook needs a real
+    reachability predicate, and manufacturing one is the over-fire trap BOSS guards against. Ties to
+    IDEA-041 (reachable → discoverable: *"does anyone know it exists?"*).
+
+## 0.91.0 — 2026-06-20
+
+- **Team wayfinding — close the gap between the shipped founder layer and the docs (IDEA-037 / FEAT-021
+  follow-on).** FEAT-021 shipped 12 releases of team functionality (v0.74→v0.85: `boss team`, `/decide`,
+  `/practice`, `mentor-cofounder`, the `coordination` moment, the state cut) but the *wayfinding* lagged —
+  exactly the README-drift pattern IDEA-018/035 exist to catch. Closed it three ways: (1) **`boss map` now
+  lists `boss team`** in its standing controls, so the command is discoverable in a live install (it was
+  invisible there before); (2) **new [`docs/GUIDE-teams.md`](../docs/GUIDE-teams.md)** — a founder-facing
+  guide to working with a cofounder: how to add a team, the loop with two of you (`/decide` · `/practice` ·
+  `boss board --mine` · `mentor-cofounder`), what the `coordination` conscience moment watches, **what's
+  shared vs. private** (push = backup + keep-in-the-loop; the conscience's relationship notes stay yours),
+  the lightweight working agreement (Driver/Approver + consent), and the hard lines (never arbitrates,
+  never scores equity); (3) **`docs/GUIDE.md` + `README.md`** gained a "Building with a cofounder?" pointer
+  + `mentor-cofounder` in the mentor table / advisor list. Docs + one `src/map.js` line; eval gate **120/0**;
+  `/tmp`-verified (`boss map` shows `boss team`). All dormant-solo framing preserved (solo founders see no
+  team noise).
+
+## 0.90.0 — 2026-06-20
+
+- **The rest of the `/vet` routing sweep — reliability, generative-UI/memory, anti-sameness, calibrated
+  pitch (RVW-040/043 + RVW-051/052 + RVW-037 → `/boss-learn` UP).** Final batch of the queued ADAPT/ADOPT
+  sweep, three surfaces, all judgment-aids — no new gates or hook predicates. (Bundled into one release to
+  minimize version churn during the concurrent FEAT-023 practice stream.)
+  - **`/evals` — correctness ≠ safety + pass^k (RVW-040/043, L1).** New "Correctness ≠ safety — the
+    adversarial half" section: a clean `/evals` pass isn't *done* until `/red-team` runs the adversarial
+    half (safety degrades under jailbreak across every model — Stanford HAI AI Index 2026; OWASP-Agentic is
+    what to probe). Sharpening gains **pass^k** (τ-bench): non-deterministic ≠ run-once — run each
+    load-bearing case k times, count how often it *all* succeeds; zero-dependency, a loop around the case
+    you already wrote. Two new Rules lines.
+  - **`ai-ux-patterns` — two net-new patterns (RVW-051, practice).** **§9 generative-UI control spectrum**
+    (static → declarative → open-ended; open-ended sits in the §4 irreversibility tier — an injected prompt
+    can redraw what the user sees) and **§10 memory-as-a-reviewable-object** (view/edit/correct/delete/scope
+    — the footprints principle extended from *what the agent did* to *what the system knows*; BOSS dogfoods
+    it as a memory-carrying tool).
+  - **`design-system` — AI-default = indistinguishable (RVW-052, practice).** Sharpened the "aesthetic
+    ambition" section: AI-default isn't just generic, it's *indistinguishable from every competitor* (the
+    Tailwind `bg-indigo-500` apology as the hook); *build faster ≠ build sameness*; **spend the time the AI
+    saved on the ~5% that's yours** — pointed at the existing distinctiveness pass, no new mechanism.
+  - **`mentor-pitch` — calibrate claims to evidence (RVW-037, L2/V1).** One attributed heuristic:
+    overclaiming *measurably lowers what founders raise* (HBR 2025); match the verb to the proof
+    (shows/suggests/we believe/we're testing). Calibration, not suppression — tied to the existing anti-hype
+    voice. The `/vet` routing sweep is now complete; remaining verdicts are NOT-YET watches (034/055/035).
+  - *Pulled via `/boss-sync`; `/evals`+`mentor-pitch` arrive at MVP/V1 unlock, practices are BOSS-canonical.*
+
+## 0.89.0 — 2026-06-20
+
+- **`scalable-architecture` practice — architecture that survives the climb (FEAT-023 thread 2,
+  `/boss-learn` UP).** Second slice of the AI-native build-process track (solo-applicable). The spine:
+  *the value is never the rule — it's the **enforcement loop**.* An agent re-derives a codebase's
+  conventions every session, so anything you only *wrote down* drifts and anything you *encoded as a check*
+  compounds (Factory.ai: "documented conventions rot; enforced conventions compound"). Two moves — **defer
+  the architecture tax you can defer, encode the conventions you can't afford to lose**:
+  - **Modular-monolith-first, extract when forced** (Fowler MonolithFirst; Shopify's 2.8M-line modular
+    monolith [EVIDENCE]; microservice envy on Hold). The *modular* half is load-bearing — clear module
+    seams inside one deployable so the eventual extraction is cheap; a `FEAT` is a natural module.
+  - **Spend rigor on the one-way doors — the schema is the one.** Migrations-as-code from the first table
+    (Bezos doors + evodb), the single thing genuinely expensive to retrofit. **The migration log is also
+    the guardrail against AI schema drift** — wire schema changes into git-workflow's high-risk review tier.
+  - **Conventions as code, enforced not remembered** — formatting-as-law (Biome's single binary is
+    agent-friendly), boundaries-as-lint + architectural fitness functions (principle + writable lint now;
+    heavy custom-plugin tooling NOT-YET), strict types at the boundary.
+  - **The ratchet holds the line, not the reviewer** — extends the existing
+    [`quality-ratchet`](../library/practices/quality-ratchet.md) (no-new-violations baseline gated in
+    `/smoke`), pointed at an architecture metric; doesn't restate it. Plus a pointer to
+    [`context-discipline`](../library/practices/context-discipline.md) for the one-canonical-context-file
+    finding (failure mode is over-length, not under-spec).
+  - **UP-only this slice** (no template DOWN): the practice is the deliverable and mentors cite practices by
+    reference; the natural `mentor-architect` DOWN is **deferred to avoid a live edit collision** with a
+    concurrent session, not on judgment. Altitude/JIT held — modular seams at MVP, migrations at the first
+    table, conventions-as-code at the second author (a second *agent* counts), services at V1→Scale. Zero-dep;
+    library only; eval gate **120/0**. **FEAT-023 thread 3 (V1→Scale org rung) remains — deferred until a
+    real V1-stage project.**
+
+## 0.88.0 — 2026-06-20
+
+- **`git-workflow` practice — trunk-based, review-bounded AI building (FEAT-023 thread 1, `/boss-learn` UP
+  + a DOWN into L1/MVP).** The first slice of the AI-native build-process track (spun out of the
+  founding-teams research; **solo-applicable**, not team-specific). The reframe: AI didn't change what good
+  version control is — it changed which part *hurts*. The agent writes code ~4× faster; **review** is the
+  new bottleneck (~12% delivered), so the whole discipline reorients around *how much two humans can stand
+  behind in a day.*
+  - **UP — new [`library/practices/git-workflow.md`](../library/practices/git-workflow.md).** Trunk-based
+    default (DORA [EVIDENCE]: ~2.3× elite; <3 active branches, merge daily; `/smoke` is the gate that makes
+    it safe; **CI is a practice, not a platform** — a 2-person smoke check *is* its CI). **Git worktrees =
+    the AI-parallelism primitive, CAPPED at ≈2–4 = your *review* capacity, not your agent count** (you can
+    spawn ten agents; you can't read ten diffs — more agents than you can review is unreviewed code with
+    your name on the merge). **Risk-tiered review, not blanket gates** — low-risk gets a glance, high-risk
+    (auth/money/migrations/deletes/deploys/AI-paths) gets the *other* human, and **BOSS's `/smoke` +
+    `/evals` + `/red-team` ARE that high-risk tier.** **Read the test diff harder than the code** (agents
+    rewrite assertions to match broken behaviour — *did the behaviour get fixed, or the expectation
+    lowered?*). **Whoever clicks merge owns what the agent wrote** (Osmani); **mob the humans+agent on hard
+    problems** (the questioning reflex degrades with an AI pair). **Honesty anchor: METR (n=16)** —
+    experienced devs on mature repos 19% *slower* with AI while *believing* 20% faster (the perception gap;
+    opposite population to a greenfield startup). **Ownership = prompt-author intent + reviewer acceptance**
+    — stated once, shared verbatim with FEAT-021's founder layer.
+  - **DOWN — folded into the L1/MVP template** (`stages/L1-mvp/template/claude-append.md`): a tight
+    "Git workflow (trunk-based, review-bounded)" section in the MVP working rules, so a scaffolded venture
+    inherits the discipline at the rung where there's a real `main` to keep green — kept lean (the practice
+    holds the depth; CLAUDE.md bloat is the failure mode the research itself names).
+  - **Altitude/JIT held** (not a Quickstart lecture — earns its place at MVP). Zero-dep; library + template
+    only; pulled via `/boss-sync`. **FEAT-023 threads 2 (`scalable-architecture`) + 3 (V1→Scale rung)
+    remain** — thread 2 next-if-earned, thread 3 deferred until a real V1-stage project.
+
+## 0.87.0 — 2026-06-20
+
+- **The mentor-architect bundle — the jagged frontier, the model underneath, evals-as-spec (`/vet` sweep:
+  RVW-046 + RVW-041 + RVW-053 + RVW-050 → `/boss-learn` UP, MVP/L1).** Four findings sharpen the AI-native
+  architecture mentor, all as judgment-aids — no new gates, no scoring apparatus imported.
+  - **New section "the jagged frontier, and the model underneath"** — two judgments that sit above every
+    AI-native build call and **move with each model release** (the model-recalibration discipline, IDEA-014):
+    - **Inside/outside the frontier (RVW-046, Dell'Acqua/Mollick 758-consultant RCT).** AI is sharply
+      additive inside the suitable set, ~19pp *less* likely correct outside it — judge which side a task is
+      on, pick **centaur** vs. **cyborg**, and **re-ask with every model jump**. Heuristic, not law.
+    - **The 70% problem (RVW-053, Osmani + GitClear telemetry).** AI gets you ~70% (what you understand) and
+      stalls on the last 30% (what you don't) — the marker between `/prototype` (sketch freely) and
+      `/spec`/MVP (now you must understand what ships). Can't-shape-the-30% = slow down, not ship.
+    - **Non-default-model transparency (RVW-041, Stanford FMTI).** When deliberately *not* defaulting to the
+      host model, weigh transparency alongside cost/capability — a four-item "what to ask" list (data
+      provenance, known limits, deprecation/retirement policy, change cadence). Indicators-as-questions; the
+      Index's scoring left out.
+  - **Reliability bullet gains evals-as-spec (RVW-050, Husain/Shankar).** For an AI product the **eval *is*
+    the spec** — writing it drags you across the *Gulf of Specification* (loose intent vs. what the model
+    actually does); define the quality bar before building. `/evals` is the machinery; this is the judgment
+    above it. The standalone `evals-driven-development` practice stays a NOT-YET-until-MVP→V1 follow-on
+    (Quickstart founders don't need eval ceremony, Principle #2).
+  - *MVP-mode mentor (`stages/L1-mvp/`); arrives at `boss unlock mvp`; pulled via `/boss-sync`.*
+
+## 0.86.0 — 2026-06-20
+
+- **The thesis bundle — BOSS's own claim, made honest, plus two coaching blades (`/vet` sweep:
+  RVW-047 + RVW-049 + RVW-039 → `/boss-learn` UP).** Three [EVIDENCE]/[THOUGHT-LEAD] findings routed into
+  the surfaces where BOSS makes its case — no new machinery, exactly as each verdict scoped.
+  - **`PRINCIPLES.md` "why" reframe (RVW-047, Camuffo 759-firm RCT [EVIDENCE]).** The honest version of the
+    thesis: disciplined validation doesn't *guarantee a win* — it makes you **decide faster, including
+    quitting faster**. Cheap AI lowers the cost of *building*, not of *being wrong* — so the cheaper
+    building gets, the more the loop matters, not less. Named because BOSS's conscience exists to stop
+    self-fooling, and that has to include BOSS not overclaiming its own promise.
+  - **`mentor-venture` gains "the evidence you carry" (RVW-047/049/039).** Three coaching lines, judgment
+    first / cite-on-request: (1) validation buys faster decisions, not wins; (2) **the demoware test** —
+    *"if your AI product only replaces a prompt plus a copy/paste, it's a demo, not a product"* (Cagan/SVPG),
+    anchored at the `/prototype`→`/spec` graduation; (3) **the competence-gate** — AI advice *amplifies* the
+    judgment a founder already has and can harm the one least able to grade it (Otis et al. HBS RCT, 640
+    founders: high performers ~+15%, struggling ~−8%), so at a call they may not be equipped to judge, ask
+    *"are you set up to judge this answer?"* and point at who'd know.
+  - **`conscience-voicing.md` gains the competence-gate as a voicing (RVW-039).** A lens the `caution`/`drift`
+    moments can reach for — a humility prompt, rare and suggestive, never a gate, **explicitly not a new
+    hook predicate** (no signal detects over-trust; don't manufacture one). Cites Otis + CHI 2025
+    automation-bias (confidence in AI tracks *less* checking, not more).
+  - *Founder-facing, project-neutral; pulled via `/boss-sync`.* mentor-venture ships from Quickstart;
+    PRINCIPLES + the practice are BOSS-canonical, synced.
+
+## 0.85.0 — 2026-06-20
+
+- **The `coordination` conscience moment — the founding-team seam, watched via artifacts (founder layer
+  slice 5b; IDEA-037 / FEAT-021).** BOSS's first *team-aware* conscience moment, and the careful gate-piece
+  of the program. Built on the most-replicated human-AI-teaming finding: **AI accelerates each individual
+  but erodes the human-to-human seam invisibly** (Ju & Aral RCT: social/emotional comms −27% while
+  *perceived* teamwork stayed flat — so a "how's teamwork?" prompt is *proven blind*; you have to watch the
+  artifact channel). New `coordination-loop` (L1): **entry** = it's a team (`.boss/config.json` has a
+  cofounder `"handle"`) **AND** real work has happened (`docs/devlog.md` ≥3 entries); **exit** = ≥1
+  `DEC-NNN` recorded together. Open → the conscience reads a **bounded** slice (decisions dir · `boss board`
+  · `boss team`) and judges: is work flowing through one founder's agent *around* the cofounder, or is the
+  deciding just happening off-repo on a call? Fires only on a real seam; **stays silent on a quiet log**
+  (the evidence is weak-transfer — over-firing would punish a healthy team that talks on calls). **Dormant
+  by construction when solo** (no `"handle"` → unopenable); **fires at most once/session**; **serves the
+  partnership as the unit and NEVER takes a side** (surfaces the seam, never whose fault); points at
+  `/decide` + `mentor-cofounder`; mutable (`boss conscience mute coordination`). The **Red-Light** moment
+  from the ai-adoption handoff was deliberately **NOT** built as a hook (it's not predicate-gateable —
+  there's no artifact that says "about to automate a teammate's task"; forcing it = the over-fire failure
+  mode) — it lives as a framing in `mentor-cofounder` instead. **Conscience eval gate: 120/0** (113 + 7 new
+  `moment-coordination` cases — fire / dormant-solo / decided-together / too-early / no-team / empty);
+  `/tmp`-verified end-to-end (fires on the seam, goes silent once a `DEC` is recorded). Zero-dep.
+  **FEAT-021 slice 5 (mentor-the-team) now complete** (5a `mentor-cofounder` + 5b the moment). **Next:
+  slice 6 — `boss credits` + the ownership/equity moment (reshaped by the research: refuse-to-score,
+  CRediT/IBM-validated, contribution kinds, structural-trigger equity — the worst solo-test profile, build
+  the evidence substrate only and get it in front of a real team).**
+
+## 0.84.0 — 2026-06-20
+
+- **Humane harm-taxonomy as a shippable practice + the mentor internal/shipped boundary made explicit
+  (RVW-045 re-homed; `mentor-architect` verdict).** Caught while routing: RVW-045's harm-taxonomy had landed
+  in BOSS's *gitignored* `mentor-humane` agent — shipping to nobody. `mentor-architect` reframed it — the
+  humane lens is **cross-cutting**, so it belongs in a practice every mentor + the conscience can cite, not
+  inside one agent ("the agent was the wrong container, not just the wrong location").
+  - New **`library/practices/harm-taxonomy.md`** — Anthropic's 5 harm dimensions + Ada Lovelace's 4
+    relationship-harms (manipulation / dependence / anthropomorphism / overreliance), the shared vocabulary
+    the conscience, every mentor, and `/canvas` §3 reason against. Pairs with the dark-pattern checklist +
+    `/red-team --humane`.
+  - **`docs/MENTORS.md`** gains an **internal-vs-shipped boundary table** on the real axis ("artifact a
+    founder's *project* consumes" vs. "tooling for authoring BOSS itself") — the thing that would have caught
+    the misroute. Clarifies: the humane *lens* ships from Quickstart; the standalone `mentor-humane` *agent*
+    stays Scale (ethics wants a hook / cross-cutting shape, not an opt-in door).
+  - Captured, not built: **IDEA-038** (`library/` as the canonical managed-artifact shelf; templates become
+    thin manifests — deferred, to be *pulled* by the learning loop, not pushed by a routing bug) +
+    **IDEA-039** (humane lens as conscience-moments + practice vs. a standalone agent — decided by a future
+    humane-moment eval set, `tester` territory) + an architecture decision record. Practice + doc only;
+    zero CLI / dependency change.
+
+## 0.83.0 — 2026-06-20
+
+- **`/decide` gains a cheap falsifier + AI-decision provenance (founder layer slice 1 iteration; IDEA-037 /
+  FEAT-021, from the research realignment).** The single highest-leverage change the founding-teams research
+  surfaced: a decision log that only assigns blame **fails — verification cost is the bottleneck**, not
+  accountability. So `DEC-NNN` now carries:
+  - **A `## Falsifier` field — *"what would prove this wrong, and by when?"*** (*"if signups don't move by
+    July, X was wrong"*), mirrored into an optional `revisit_by:` date. **Required for `costly`/`one-way`
+    calls, encouraged for reversible** — it makes finding out you were wrong *cheap and scheduled*.
+  - **`decided_by:`** — `founder` / `ai-suggested-ratified` / `ai-autonomous`, so over-delegation of
+    load-bearing calls to the model is *visible* (where automation bias enters).
+  - **Reversibility-scaled ceremony** — reversible = a line + the **consent** question (*"is it safe enough
+    to try?"*, which gives a non-technical cofounder honest language to agree without fully evaluating a
+    technical call); `costly`/`one-way` = require the falsifier + weigh the alternatives.
+  - **A mild skeptic prompt at irreversible *AI-suggested* decisions** (*"what's the one thing you'd check
+    before you can't undo it?"*) — **a prompt, never a gate** (forced verification *backfires*; disposition
+    beats process). BOSS's own `DEC-001` updated to dogfood the new format.
+  Evidence: automation-bias review + "Bias in the Loop" (accountability-alone insufficient; forced
+  verification backfires; skepticism beats every design factor), "Don't Vibe" (over-delegation seam),
+  sociocratic consent. Skill-layer only, zero-dep; eval gate **113/0**; `/tmp`-verified. **Also triaged the
+  research:** FEAT-021 sharpened across all slices; build-process/scaling work **spun out to FEAT-023**
+  (`git-workflow` + scalable-architecture + the V1→Scale rung — not team-specific). **Next: slice 5b (the
+  teams-conscience family — Red-Light + watch-the-seam + moral-crumple-zone, gate-evalled).**
+
+## 0.82.0 — 2026-06-20
+
+- **The humane lens, operationalized — dark-pattern checklist + `/red-team --humane` + a harm taxonomy
+  (RVW-031 + RVW-045, via `/vet` → `/boss-learn`).** Third routed bundle from the research session, turning
+  the humane lens from a vibe into named, checkable artifacts:
+  - **`library/practices/ai-ux-patterns.md`** gains a **dark-patterns** section — CDT's *Dark Patterns in AI
+    Chatbots* (2026, CC-BY) 37 patterns in 5 families (data/memory exploitation · misleading design ·
+    autonomy-for-engagement · false social/emotional connection · coercive monetization), framed as
+    *recognize-as-you-build*, plus CDT's **constructive "humane alternative"** (default conversations to
+    end, opt-in social layer, genuine delete controls, no emotional language near a purchase). Key insight:
+    these can **emerge from the model** (sycophancy), not just be designed — so test the *built* product.
+  - **`/red-team` gains a `--humane` dimension** — probes the founder's *own* AI product for the dark-pattern
+    families, weighted toward emergent ones (sycophancy, engagement-prolonging, emotional-manipulation-near-
+    money, misrepresentation). Binary pass/fail; **suggestive, never blocking** (conscience-not-censor).
+  - *(BOSS-internal, not shipped)* BOSS's own self-hosted `mentor-humane` also gained a **harm taxonomy** —
+    Anthropic's 5 harm dimensions + Ada Lovelace's 4 relationship-harms (manipulation / dependence /
+    anthropomorphism / overreliance), reflexively disciplining BOSS's own voice. RVW-045 had no founder
+    template to route into (mentor-humane is BOSS-only), so this is dogfood, not a shipped capability.
+  - Held the conscience-not-censor line throughout (names the cost + the humane alternative; never makes a
+    choice unavailable). Practice + skill + agent text only; zero CLI/dependency change.
+
+## 0.81.0 — 2026-06-20
+
+- **`mentor-cofounder` — mentor the TEAM (founder layer slice 5a; IDEA-037 / FEAT-021).** Built on the
+  founders' Pain B (*"different founding skill sets — an engineer who's a first-time CTO, or a non-tech
+  person who doesn't know how to work with an engineer cofounder on building together"*). Every other
+  mentor coaches *a* founder; this new L1 mentor coaches the **relationship between founders** — the
+  differentiated "mentor the team" value. It helps a team **divide work across skill sets** (record
+  who-owns-what as a `DEC`), **bridge the skill gap both directions**, **surface the hard conversations**
+  (roles/pace/equity — the talks that kill teams by never happening), and **name decision rights** (one
+  DRI; equal equity fine, 50-50 control-deadlock the trap). **Folds in the inherited AI-adoption-culture
+  knowledge** (from the v0.80.0 practice, the handoff source): the **Human Agency Scale / Red-Light** zone
+  (don't automate a teammate's task they don't want automated), **psychological safety paired with high
+  standards** (Edmondson / RVW-035 — not niceness), killing the **secret-cyborg** dynamic (reward honest AI
+  use, shared via `/practice`), and the **no-workslop norm** (*"would I be proud to hand this to my
+  cofounder?"*). **Wires the AI consent + norms conversation into the `boss team` flow** — a one-time nudge
+  on the solo→team transition + a standing pointer in the team view. **The bright line is built in:** it
+  serves the **partnership-as-the-unit** and **NEVER takes a side** between cofounders (facilitate, name the
+  tradeoff, they decide; the only override is harm, not preference; not a lawyer or cap table). **Dormant
+  when solo.** Registered in the L1 manifest + `docs/MENTORS.md` (arrives at MVP). Zero-dep; eval gate
+  **113/0**; `/tmp`-verified. **Next: slice 5b — the conscience Red-Light *moment* (new hook detector +
+  eval cases through the gate; the deliberate, careful piece) · then slice 6 (`boss credits` + the
+  ownership moment).**
+
+## 0.80.0 — 2026-06-20
+
+- **New practice: `ai-adoption-culture.md` — bring AI to a team without breeding resentment (RVW-038,
+  via `/vet` → `/boss-learn`; feeds IDEA-037).** The second routed bundle from the research-feeds session,
+  and the most net-new: how a small founding team adopts AI so people **opt in** rather than comply.
+  Distilled from Stanford's **Human Agency Scale** (automate what's *wanted*; name the "Red Light"
+  capable-but-unwanted zone), **Edmondson** psychological safety (make it safe to admit "I don't know AI";
+  pair safety with high standards), Mollick's **"secret cyborgs"** (reward honest AI use, don't punish it
+  into hiding), and the **"workslop"** finding (sloppy AI output erodes trust *between cofounders* — own
+  your draft before you forward it). Ships a concrete **cofounder AI consent + norms conversation** and
+  stays JIT — dormant solo, surfaces when a second person joins, **never a mandate** (the conscience names
+  the Red-Light tension, never picks the answer). **Authored as a reviewable starting draft for the
+  IDEA-037 founding-teams build** — the handoff names the recommended wiring (a conscience Red-Light
+  moment, the consent conversation in `boss team`, mentor-the-team citing it) and leaves the hook/gate
+  changes to that build. Library practice, zero-dep (not scaffolded); folds in RVW-035 (Edmondson) when
+  slice 5 lands.
+
+## 0.79.0 — 2026-06-20
+
+- **Agent-security hardened for the 2026 agent-native surface (RVW-032/042/044/054, via `/vet` →
+  `/boss-learn`).** The first routed bundle from a research-feeds mining session (new
+  `docs/research/SOURCES.md` — a ~40-org institutional roster — plus a 25-verdict `/vet` sweep,
+  RVW-031…055). `library/practices/agent-security.md` + the L1 `/red-team` skill grow beyond the
+  stateless lethal-trifecta / LLM-Top-10 baseline to cover what a founder shipping an *agent* actually
+  faces:
+  - **Agent-native threat model** — the **OWASP Agentic ASI Top 10 (Dec 2025)** (goal hijack, tool
+    misuse, agentic supply chain, memory/context poisoning, identity/privilege abuse, rogue agents…)
+    added to `/red-team` for any target with tools + memory + autonomy; the stateless LLM Top 10 still
+    covers a plain prompt-in/text-out path. Plus **agentic misalignment** named as a *measured* failure
+    mode (Anthropic) — bound autonomy + sensitive access; don't assume good behaviour.
+  - **Containment defaults** beneath the Rule of Two (Anthropic "how we contain Claude" + Redwood
+    control): **match isolation to oversight** — read-only / read-write-no-delete mount tiers, **egress
+    allowlists**, inspect-tool-returns-before-context, and **gate the irreversible** behind a human or a
+    cheaper trusted check.
+  - **The shipped app is its own surface** — AI defaults to insecure code (Veracode: ~45% of
+    AI-generated code ships an OWASP-Top-10 vuln) and the classic vibe-coded leak is **client-side key
+    exposure** (the Tea breach; ~25k secrets found across vibe-coded sites). A new **pre-ship
+    app-security pass** in `/red-team` (no secrets/keys in the bundle + OWASP web basics),
+    **non-negotiable for `first-product` / `vibe-coder-newbie`**. Names the honest gap: **`secrets-guard`
+    does NOT cover this** — it stops the *agent* reading secrets into context; it does nothing about a
+    *shipped app* exposing a key.
+  - Held the JIT line (Principle #2): surfaces one trigger at a time — trifecta on first untrusted read,
+    the ASI list on first agent ship, the pre-ship scan on first deploy, the full battery for regulated
+    cohorts. Practice-doc + skill-text only (zero CLI / dependency change); provenance carries the RVW
+    trail.
+
+## 0.78.0 — 2026-06-20
+
+- **`/practice` + `PRAC-NNN` — the shared craft commons (founder layer slice 4; IDEA-037 / FEAT-021).**
+  The most *differentiated* slice, built on the founders' Pain C: *"either of us could be discovering best
+  new ways to use agentic AI — it changes so fast — and we want to keep sharing + staying current, using
+  best practices, so we can focus on building, with mentorship so we're not worrying whether we're outdated
+  or expensive."* New L1 skill `/practice` captures a **craft learning** (a better/cheaper/newer way to
+  build with AI) as a shared, attributed `docs/practices/PRAC-NNN-<slug>.md` (What we learned / Why it
+  works / How to apply), stamped with **who learned it** (`@github-username`) so a teammate gets it *and*
+  knows who to ask. **Shared by construction** (commits with the repo per the slice-3a cut — push backs it
+  up, a cofounder who clones has it). **Staleness-aware — the part that keeps you *current*, not just
+  documented:** a practice tied to a specific model/price/tool can carry `review_by:`, and when it passes,
+  `/revalidate` asks *"still the best way? anything changed?"* → keep / update / retire — the team's quiet
+  defense against being outdated or overspending, the model-recalibration discipline ([[IDEA-014]])
+  team-scoped. **Held the humane line:** attribution is recognition + a pointer, **never a scoreboard**
+  ("measures what the team knows, never who contributed more"). New `PRAC-NNN` ID type (`docs/IDS.md`);
+  registered in the L1 manifest. Zero-dep, skill-layer only; eval gate **113/0**; `/tmp`-verified (ships,
+  `boss map` lists it, `PRAC-001` parses with attribution). **Next (FEAT-021): slice 5 mentor-the-team ·
+  slice 6 `boss credits` + the ownership moment.**
+
+## 0.77.0 — 2026-06-20
+
+- **The founder-layer state cut — back up + share, keep the conscience private (slice 3a; IDEA-037 /
+  FEAT-021 / DEC-001).** Occasioned by Ajesh's insight that BOSS taught us the *released-app vs.
+  dev-codebase* split: BOSS gitignores its *own* docs (don't ship private strategy in a public OSS package),
+  but a **venture's `docs/` already commit** — so a scaffolded team's ideas, canvas, `DEC-NNN` decisions,
+  research, RESUME, and the venture brain's `read.md` are **already backed up (push = backup) and shared
+  (a cofounder who clones is in the loop)**. This release names that win and **plugs the one leak**: the
+  template now gitignores the **per-person** conscience state — `.boss/brain/relationship.md` (what the
+  conscience said to *you* and what you did with it) + `.boss/trace.jsonl` — so one founder's private nudge
+  history never travels to the other (Contextual Integrity). The backup/share framing is surfaced in
+  `/welcome` (first-run) and the L0 `AGENTS.md` conventions. **The decision is recorded as BOSS's own first
+  `/decide`** — [`DEC-001`](../docs/decisions/DEC-001-founder-layer-brain-cut.md): *venture brain `read.md`
+  SHARED (the hive-mind read, also seeds a joining cofounder's brain for free), conscience `relationship.md`
+  PER-PERSON* — drafted with the research- + 3-mentor-backed recommendation, **Ajesh to confirm or supersede**.
+  Reversible at a cost (the leak fix *prevents* the irreversible accident; sharing `read.md` confirms the
+  status quo). Deferred: per-founder namespacing (`.boss/founders/<handle>/`) — the structural fix, needs
+  `brain.js` path changes. Template-only + docs; `/tmp`-verified (per-person ignored, shared/backup commit).
+  **Next (FEAT-021): slice 4 — shared craft commons (Pain C, the most differentiated) · slice 5
+  mentor-the-team · slice 6 `boss credits` + the ownership moment.**
+
+## 0.76.0 — 2026-06-20
+
+- **`boss board` owner lens — `owner:`-as-person + `--mine` (founder layer slice 2b; IDEA-037 / FEAT-021).**
+  The board (already a pure projection of frontmatter) now reads `owner:` and, **when the venture is a
+  team**, shows the founder who owns each card (`@handle`) — provenance of *who's the DRI*, surfaced as a
+  quiet suffix. `boss board --mine` narrows to the cards you own ("what am I on the hook for"); the JSON
+  projection (`--json`) carries `owner` per card for agent-readability. **Dormant-solo:** a solo founder
+  sees no owners and nothing changes — the lens only lights up once `boss team` has a cofounder. **Held the
+  humane line:** only a `@handle` counts (role owners like `pm` are ignored), and owners are shown as
+  per-card provenance, **never aggregated into a per-person count/leaderboard** (the credit-score line
+  mentor-humane drew). Quote-tolerant (`owner: "@handle"` — a leading `@` is reserved in YAML). Pure
+  projection preserved (no new state); zero-dep; eval gate **113/0**; `/tmp`-verified (team-shown /
+  solo-hidden / `--mine` filter / JSON field). *Deferred:* owner in the HTML board + `--next`/`--blocked`
+  views. **Next (FEAT-021): slice 3 — keep-in-the-loop + the shared/personal state cut** (the one
+  costly-to-reverse decision; to be recorded as BOSS's own `DEC` once the cut is chosen).
+
+## 0.75.0 — 2026-06-20
+
+- **`boss team` — the team-aware foundation (founder layer slice 2; IDEA-037 / FEAT-021).** Makes BOSS
+  *know* whether a venture is solo or a founding team, keyed on **GitHub identity** (`@username` via
+  `gh api user` → `git config`, never fabricated — the principal id the whole team layer builds on).
+  `boss team` shows the venture's people; `boss team add @handle "Name"` / `remove @handle` manage the
+  roster (stored in `.boss/config.json`). **Dormant-solo by design** (the mentor-pass guardrail that
+  clears the solo test): with no cofounder declared, a solo founder sees nothing new and *nothing
+  changes* — the team layer only lights up when someone joins, so it's inert, never overhead. Guards
+  against adding yourself (you're already "you"). `/welcome` now asks the light, optional "solo or with
+  someone?" question at first run (flippable anytime); L0 `CLAUDE.md`/`AGENTS.md` wayfinding refreshed for
+  `/decide` + `DEC-NNN` + `boss team` (no drift). `src/team.js`, zero-dep; conscience eval gate **113/0**;
+  `/tmp`-verified (solo→add→team→remove + self-add rejection). **Where the roster lives is deliberately
+  LOCAL for now** — whether it should travel via git (so a cofounder sees it) is the shared-vs-personal
+  **state cut**, slice 3, which will be recorded as BOSS's own `DEC` (dogfooding `/decide`). **Next
+  (FEAT-021):** slice 2b `owner:`-as-person + board-by-owner · slice 3 keep-in-the-loop + the state cut.
+
+## 0.74.0 — 2026-06-20
+
+- **`/decide` + `DEC-NNN` — the decision log (founder layer, slice 1; IDEA-037 → FEAT-021).** First build
+  of "BOSS for founding teams," green-lit on **real-founder demand** (past-pain stories from founder
+  conversations — the first field signal against the canvas's n=0 demand risk, Risk #6 antidote). New L0
+  skill `/decide` records a **load-bearing or hard-to-reverse** choice as a durable ADR-lite record in
+  `docs/decisions/DEC-NNN-<slug>.md`: Context / Decision / **Why** / Consequences, stamped with the
+  **decider** (`@github-username`, resolved from `gh api user` → `git config user.name`, never fabricated)
+  and a **`reversibility:`** flag (`reversible | costly | one-way`, Bezos two-way/one-way doors).
+  **Supersede-don't-edit** — a changed mind writes a new `DEC` with `supersedes:` and flips the old to
+  `status: superseded`, so the chain *is* the story of how thinking evolved. The rationale future-you (and
+  a cofounder who wasn't in the room) can read instead of guessing — and in a team, the artifact both can
+  point at instead of misremembering. New `DEC-NNN` ID type in `docs/IDS.md`; registered in the L0
+  manifest (decisions happen from day one). **Bright lines held** (mentor-vetted, IDEA-037): records a
+  decision, never gates one; never a cap table or legal advice (points at a real attorney for
+  equity/vesting); the conscience may surface a tension but **never picks a side between cofounders**.
+  Zero-dep, skill-layer only (no `src/` change). **Slice 1 of a 6-slice program** (team-aware foundation →
+  keep-in-the-loop → shared craft commons → mentor-the-team → credit + the ownership moment) — each later
+  slice records *into* this one. See [`FEAT-021`](../docs/ideas/FEAT-021-founder-layer-decision-log.md).
+
+## 0.73.0 — 2026-06-20
+
+- **`mentor-business` (V1) gains the on-ramp + tier-design layers — the founder's pricing menu, completed
+  (RVW-030, ADAPT).** RVW-023 (v0.66) gave the V1 business mentor its *metering-basis* axis (per-seat →
+  usage → hybrid → outcome). This adds the two questions that axis left open, sourced from a `/deep-research`
+  pass (27 sources, 22/25 claims adversarially confirmed; `docs/research/pricing-and-tiers-playbook-2026.md`):
+  - **The on-ramp** — freemium / free-trial / reverse-trial / no-free-tier, chosen by *available traffic*
+    and *per-user cost*, with the AI catch made explicit (every free interaction burns real compute, so
+    freemium needs cheap/hard-capped free or a reverse trial instead of a crippled tier).
+  - **The tiers** — the ~3-tiers-plus-enterprise default, gating each tier on the axis matching the value
+    metric and *only where a tier's absence blocks a segment's core job*, free→paid line at a real aha.
+  - Carries three cautionary cases (Cursor's credit-repricing apology, "customers don't think in tokens,"
+    vague outcome-units backlash) and an explicit **numbers-are-contested** rule: coach the ordering and
+    trade-offs, never quote a precise conversion rate as fact.
+  - **Template only**, not BOSS's own `mentor-business` instance — BOSS has zero paying users, so its own
+    agent stays deliberately lean (Principle #2). Inherits the agent's existing defer-discipline and
+    "voice the tension, never filter the menu" humane rule. Connected projects pull it via `/boss-sync`.
+
+## 0.72.0 — 2026-06-20
+
+- **Per-moment mute + first-run consent — "don't voice it if I don't want it," at the granularity of
+  the moment (v0.71.0 conscience-voicing follow-on).** Pause (v0.23) silenced the *whole* conscience for
+  a bounded session; this adds the surgical half the founder can't get from pause:
+  - **`boss conscience mute <moment> [--for 7d | --until-resume] [--reason]`** + **`unmute <moment>`
+    / `--all`** — turn down ONE moment (drift, caution, capture, focus…) while the rest keep speaking.
+    Hook-enforced (the conscience filters muted signals after detection, then exits silent if nothing's
+    left — same shape as pause). **Auto-unmutes on expiry**, the per-moment twin of pause's silent
+    auto-resume. Stored under a separate `conscienceMutes` key in `.boss/config.json` so pause/resume
+    (which overwrite `cfg.conscience`) can never clobber a mute — the two controls are orthogonal by
+    construction. Moment names validate against the project's actual loops (typo → the available list).
+  - **`boss conscience status` now surfaces live mutes** (like it surfaces an active pause), so a
+    forgotten mute can't silently swallow a moment forever; the over-fire-smell hint in
+    `boss conscience activity` now points at mute as the surgical alternative to pause.
+  - **First-run consent moment** — `/welcome` now introduces the moments as a set and names all three
+    controls (pause / mute / override) *before* any fires, cohort-aware (full tour gets the three-control
+    walk; the 30-second version gets pause + mute in one breath). The founder meets the conscience and
+    learns they can dial each moment, rather than discovering the controls only after being nudged.
+  - This operationalizes the `conscience-voicing` practice's consent boundary: all current hook moments
+    are *self-regarding* (about the founder's own venture discipline), so a flat per-moment mute is the
+    correct, fully-honored control. Zero-dep; single source (the L0 hook + runtime feed both the hook and
+    the `boss conscience` CLI).
+
+## 0.71.0 — 2026-06-20
+
+- **Conscience voicing — name the tension, never filter the menu.** Closes a paternalism seam found
+  while auditing what BOSS recommends to founders: the conscience is *suggestive*, but in a few places
+  it had drifted toward *withholding*. Three moves, one principle (a conscience makes a cost **visible**;
+  a censor makes a choice **unavailable**):
+  - **`mentor-business` (V1 template) gained a second axis.** The model menu was structure/licensing
+    only (OSS, open-core, patronage, cohort, SaaS…); it now also names the **metering basis** the AI era
+    runs on — per-seat → usage → hybrid → outcome → service-as-software → agent-to-agent — each with its
+    humane tension as an *overridable note*. Founders see the full menu including the models BOSS is wary
+    of; omitting them "to protect" is itself a dignity cost. New rule: **voice the tension once, then
+    yield.** (Re-opens **RVW-023** NOT-YET → **ADAPT**: per-seat assumptions are now demonstrably stale —
+    Intercom/Zendesk/SAP/Adobe live — and the verdict had conflated *should BOSS adopt this for itself*
+    (still NOT-YET) with *should BOSS tell founders it exists* (yes).)
+  - **`mentor-humane` override authority clarified — over other mentors, never over the founder.** The
+    "lens is non-negotiable" language was the one place the conscience read like a censor. Reframed: the
+    lens makes a harm *un-ignorable*, not a choice *unavailable*. Encodes the consent boundary —
+    **self-regarding** tension (the founder's own venture) is fully muteable; **third-party harm**
+    (someone not in the room) is named once even if unwelcome, because the harmed party never consented
+    to being silenced. Always *name*; never *override*.
+  - **New library practice `conscience-voicing.md`** (UP via `/boss-learn`) — the inheritable spine:
+    the conscience-vs-censor line, the 7-habit craft of voicing concern without blocking, the
+    consent-boundary table, where it applies (hook moments, every mentor, `/vet`, any menu), and the
+    existing machinery (`conscience pause`, `relationship.md`) to build on rather than reinvent.
+
+## 0.70.0 — 2026-06-20
+
+- **Wayfinding-drift check — the de-rot pass becomes a standing guard (IDEA-035, built).** v0.68 was a
+  *manual* catch-up after the hand-authored prose lagged 20 releases of new capability. The generated
+  surfaces (`boss map`, `CHEATSHEET.md`, `SKILLS.md`) can't rot — they're rebuilt from `src/modes.js`;
+  the curated prose can, and nothing flagged it. New dev-only `scripts/check-wayfinding-drift.js`
+  (`npm run check:wayfinding`, and a courtesy nudge at the end of `gen:docs`) greps `GUIDE.md` — the one
+  prose doc *meant* to walk the whole ladder — against the manifest skill lists and warns on any skill
+  named in **no** rung. **The trap, designed around:** README/`/welcome` deliberately don't enumerate
+  skills (they point at `boss map`), so the check guards *only* `GUIDE.md`, never blanket coverage —
+  and it **nudges, never blocks** (exit 0 always; a drift check that fails a commit is the ceremony BOSS
+  refuses). Internal/meta skills carry a *printed* exempt list (`boss-learn`, `design-tokens-init`,
+  `extract`, `drift-deep`), never silent. Cleared the drift it found on first run: `GUIDE.md` now names
+  `/import` + `/cost-review` in their rungs and `/boss-sync` + `/feedback` as standing utilities.
+  Dev-only (not shipped in the package); the founder-facing `boss sync` generalization stays a NOT-YET
+  UP candidate (PRINCIPLE #2). Zero-dep.
+
+## 0.69.0 — 2026-06-20
+
+- **`shipped_on:` — a true date-windowed Shipped archive (IDEA-034 follow-on).** The board's Shipped
+  column was bounded only by a recent-*count* cap (v0.66). This adds the *date* half the founder asked
+  for: stamp `shipped_on: <date>` when a FEAT ships (in `/spec`'s lifecycle note, alongside
+  `building_since:`), and `boss board` folds any ship older than ~30 days into the "+N shipped earlier"
+  `<details>` — so the column shows what landed *lately*, not every ship forever. **Frontmatter-true,
+  never guessed:** no `shipped_on:` → graceful fallback to the count cap (legacy ships still bounded).
+  Shipped now sorts newest-first by ship date (dated ahead of undated, then id). `--all` still reveals
+  everything; `--json` carries `archived` + `shippedAgeDays`. Zero-dep; verified in `/tmp` (a 30+-day
+  ship archives, recent ones show, undated legacy falls back to the cap; HTML folds the same).
+
+## 0.68.0 — 2026-06-20
+
+- **De-rot pass — the hand-authored wayfinding caught up to 20 releases of new capability (IDEA-018).**
+  The generated surfaces (`boss map`, `CHEATSHEET.md`, `SKILLS.md`) stayed current automatically; the
+  *prose* a founder actually reads had drifted — the recurrence of the exact "19-release README drift"
+  IDEA-018 was built to catch. Surgical, scope-correct fixes (each surface names what fits *its* level,
+  and still points at `boss map` for the live list — no re-enumerating):
+  - **README** — the install flow now shows `/prototype` ("hit go") and, crucially, **`boss adopt`** for
+    an already-started repo (a stranger would otherwise read BOSS as greenfield-only — the exact gap
+    IDEA-005 closed). `--ai` mentioned.
+  - **`/welcome`** (the first-run orientation) — `/prototype` is now offered as a third path ("if you'd
+    rather *see* the idea than describe it"), and `/persona` is in the skill list.
+  - **L0 `CLAUDE.md`** — `/comprehend` added to the skills line.
+  - **`docs/GUIDE.md`** — the rung-by-rung walk now includes `/prototype` + `/persona` (Quickstart),
+    `/revalidate` + `/judge-traces` + `/red-team` + `/consult` (MVP), and a new "you already started —
+    there's a repo" entry for `boss adopt`.
+  No code change; eval 113/0. The lesson re-learned: generated wayfinding is drift-proof; curated prose
+  needs a deliberate de-rot pass after a big build run — this was it.
+
+## 0.67.0 — 2026-06-20
+
+- **The `/vet` sweep's six ADAPTs, routed via `/boss-learn` (RVW-015→026).** A 12-claim skeptical sweep
+  of the research inbox (2 REJECT · 4 NOT-YET · 6 ADAPT · 0 ADOPT — a good skeptic's spread) produced six
+  scoped adaptations, now landed. The headline two sharpen what just shipped:
+  - **Outcome ledger (RVW-021) — the humane alternative to a notification cap.** `boss conscience
+    activity` now reads `.boss/brain/relationship.md` and reports an **acted-on rate** ("75% of nudges
+    landed or were engaged · N landed · N overrode · N ignored"). A persistently-low acted-on rate is the
+    *real* over-fire smell — better than a raw count, and it **never muzzles a load-bearing warning** the
+    way an arbitrary daily cap would (BOSS's fires are predicate-earned, not engagement push). **Completes
+    IDEA-013's deferred self-throttle by outcome, not by fiat.**
+  - **Brain staleness is a write-side job (RVW-026).** `/close`'s brain-write step now re-checks the
+    standing summary *before* appending and **revises/retires stale claims** — "the brain evolves, it
+    doesn't just accrete; the most dangerous brain cites yesterday's truth with today's confidence."
+    Hardens the v0.65 venture brain.
+  - Four lighter skill sharpenings: **`/prototype`** gains the build-to-learn / build-to-earn frame
+    (Cagan; RVW-016); **`/canvas`** gains a scoped humane *build-or-buy?* cell for tool-shaped ideas
+    (Fried; RVW-018); **`/spec`** gains a delegation line — *what will you verify + what's out of the
+    agent's authority* (Mollick's checklist kernel, minus the "know what good looks like" platitude;
+    RVW-020); **`/evals`** gains an AISI-Inspect pointer for trajectory eval (RVW-025; the principle was
+    already there). The 2 REJECTs (AI-runs-your-interviews → fails #6; single-strong-agent → confirms
+    IDEA-028) and 4 NOT-YETs (constrained-decoding, MCP-publishing, OTel-GenAI, outcome-pricing) are
+    recorded with re-open conditions. eval **113/0** + GRADED 24 (the conscience change is to the
+    *activity readout*, not the hook firing). Research hygiene: inbox cleared to a clean drop-zone, 12
+    new verdicts written, sources archived to `reviewed/`.
+
+## 0.66.0 — 2026-06-20
+
+- **Board intelligence — the board stops being a mirror and becomes something you (and the agent)
+  steer by (IDEA-034).** The board was already ahead of most "AI board" advice — it's a *pure
+  projection* of `status:` frontmatter, never a maintained doc (IDEA-015). So this pass doesn't add
+  richer kanban (drag-drop, swimlanes, story points, a `board.json` are all **refused** — each
+  reintroduces a second source of truth or premature ceremony); it makes the projection answer harder
+  questions and feed the conscience. Four tracks:
+  - **A — Agent-readable board.** `boss board --next` (an ordered "what to pick up," finish-before-you-
+    start), `--blocked` (everything not moving — blocked + aging + review-due, in one place), and
+    `--json` (the machine-readable projection — the actual agent-readability contract; the agent reads
+    the board as task-queue instead of re-deriving state). CLI-level, every mode; a lighter cousin of
+    the V1 `/board` skill.
+  - **B — Time-in-build aging + a bounded Shipped column.** A FEAT that's sat in Building past ~3 weeks
+    now flags `⌛ Nw in build` (the zombie-feature smell `/revalidate` targets) — **frontmatter-true,
+    never guessed**: `/spec` stamps `building_since:` when it sets `status: building`; no date → no
+    flag. Building sorts longest-running-first (finish what's been open longest). The otherwise-unbounded
+    Shipped column caps to the most recent few (`--all` / a `<details>` expander to see the rest).
+  - **C — Honest flow in `boss insights`.** Idea→build cycle time (median, from recorded `created:`
+    dates only — omitted, never guessed, when absent). Loop-closure cycle time, **never throughput /
+    velocity** (the vanity metric BOSS refuses to expose).
+  - **D — Board → conscience `focus` moment.** ≥4 FEATs in Building with nothing Shipped opens a new
+    `focus-loop` (L1-mvp): the "stop starting, start finishing" smell. Judge-style (the model reads the
+    board and distinguishes scattered abandonment from honest parallel work before voicing), at most
+    once per session, **never a gate** — and it auto-silences the moment anything ships (exit = ≥1
+    shipped). Conservative threshold + auto-silence are the over-fire guards.
+  - **E — Lightweight priority.** Optional `priority: high` in FEAT/IDEA frontmatter floats a card to
+    the top of its column (a `⬆` marker) and leads `--next` — **one level, frontmatter-true, never a
+    drag-to-reorder.** Deliberately no P0/P1/P2 ladder (that turns the board into a planning surface you
+    tend instead of ship); the caveat — *re-prioritizing isn't progress; finishing is* — ships in
+    `/spec`. Priority is the one explicit ordering signal layered over the default finish-first sort.
+  - **HTML kanban — a real visual refresh, then a sharpening pass** (the founder asked if it could
+    "feel more legit," then for stronger hierarchy): owned accent + signature dot + uppercase kicker,
+    count pills, card depth/hover, **bold titles with a quiet monospace id**, **tinted backgrounds so
+    aging/blocked/review cards pull the eye first**, a priority pill, the Shipped `<details>` fold —
+    calm and crafted, still zero-dep / single-file / light-dark (not a startup-bro dashboard).
+  - **Discipline held:** zero-dep; the model owns judgment, the CLI owns the projection. Conscience gate
+    **113/0** (105 + 8 new `moment-focus` cases) and the 24 GRADED judgment evals stay green; verified
+    end-to-end in `/tmp`.
+
+## 0.65.0 — 2026-06-20
+
+- **`relationship.md` — the venture brain's missing half: the conscience learns whether its nudges
+  land (IDEA-022 / FEAT-022).** The architecture designed two model-owned files — `read.md` (the POV,
+  shipped) **and** `relationship.md` (what the conscience *said* and what the founder *did* with it).
+  Only the read existed; this builds the relationship log, which closes the loop the frequency ledger
+  (IDEA-013) only *counts*: did the nudge **land**?
+  - **`/close` writes it** — *only when the conscience actually fired this session* — a dated entry:
+    what was flagged + what the founder did, tagged honestly (*landed* / *ignored* / *overrode, with the
+    reason* / *pushed back and was right* — the last being the most valuable: it's how the conscience
+    learns to fire better). The must-nots carry over: it logs the conscience's *own* hit rate, never
+    scores the founder.
+  - **The conscience reads it back to calibrate (the payoff).** When a moment fires, the hook hands the
+    model a **bounded** slice of the recent log (last ~1-2 sessions) so it *adjusts instead of repeats*:
+    if it's raised a point and the founder moved past it for a stated reason, it says it lighter or stays
+    silent; if a past nudge landed, it builds on it. This is what makes the conscience feel like it
+    *remembers the conversation*, not just the venture.
+  - **CLI:** `boss brain --relationship` (view the log), `boss brain record --kind relationship`
+    (stamp it; the index now carries `kind` to distinguish read vs relationship), `boss brain` surfaces
+    a one-line pointer, and `boss brain forget --before <date>` prunes **both** files symmetrically
+    (living memory across both).
+  - **Cost + safety held:** read only when a moment is firing, bounded (~900 chars), **byte-identical
+    when no log exists** → 105 gate + 24 GRADED judgment evals stay green (verified). Zero-dep; the
+    model owns the prose, the CLI owns the index. **FEAT-022 (the venture brain) is now complete —
+    read + relationship + index + living memory.**
+
+## 0.64.0 — 2026-06-20
+
+- **AI-native scaffolder — `--ai` + `/comprehend` (IDEA-022 Track 3; the last track, most guarded).**
+  Scaffold from what BOSS *understands*, not just a fixed template copy — built exactly to the
+  guardrail: **additive, behind a flag, the deterministic template stays the default.** `boss new --ai`
+  / `boss adopt --ai` do the normal reversible scaffold, then set `aiNative` in `.boss/config.json` and
+  point at a new model-driven **`/comprehend`** skill (L0). The CLI never calls a model (zero-dep, layer
+  1) — the comprehension is the skill's job (same predicate/runner split as `/import`). `/comprehend`
+  reads what BOSS can honestly understand (an adopted repo with the wide context · the captured idea +
+  `docs/source/` · or nothing-yet → it says so and stops), then **non-destructively**: fills the
+  `AGENTS.md` overview with a real read, **seeds the venture brain** with an honest first dated read
+  (so the conscience has continuity from day one — connects Track 3 → Tracks 0/4) + stamps the index,
+  and **recommends** (never auto-applies) the disciplines that fit (`/ai-first-init`, `/design-tokens-init`,
+  `secrets-guard`+`/red-team`, `/persona`). **The guardrail is in the skill itself:** everything is a
+  plain-text, diffable, revertable write in the working tree — *"if it can't be diffed, it doesn't
+  ship"*; it never rewrites the deterministic scaffold. Verified: `--ai` sets the flag + surfaces
+  `/comprehend` (and plain `boss new` is unchanged, `aiNative:false`); eval 105/0, GRADED 24.
+  **IDEA-022 is now complete — all four tracks + the spine shipped.** (The fuller presence/identity
+  design + the proactive presence-moment stay deliberately deferred: a new unprompted trigger is the
+  over-fire risk the conscience guards against; the v0.63 voicing *is* the presence.)
+
+## 0.63.0 — 2026-06-20
+
+- **The conscience now voices *with* the venture brain (IDEA-022 Track 4 — "the brain, voiced").** The
+  spine (Track 0) gave the brain a read; this makes the conscience **speak with it**. When a moment
+  fires, the hook reads a **bounded** slice of `.boss/brain/read.md` (the standing summary + the single
+  most recent dated read — continuity, not the whole history) and hands it to the model as a
+  *Continuity* frame, so the nudge is **specific to what the conscience already understands** — the
+  "how did it know that" that earns trust ("you've rebuilt onboarding three times and still haven't
+  talked to a user") instead of a generic line. The instruction is explicit: voice *with* the read,
+  don't read it back as fact, and **trust what you see now over the brain** (the founder can correct
+  it). **Cost + safety held:** the brain is read **only when a moment is already firing** (past the
+  silent early-exit — never every prompt), bounded to ~1400 chars, and **byte-identical output when no
+  brain exists** — so the **105 gate evals + 24 GRADED judgment evals stay green** (verified; the
+  judgment fixtures have no brain, so they're unaffected). The proactive "presence moment" (the
+  conscience surfacing its read unprompted) is deliberately **not** built — a new always-on trigger is
+  the over-fire risk the conscience itself guards against; the voicing *is* the presence. **IDEA-022:
+  Tracks 0, 1, 2, 4 now shipped; Track 3 (AI-native scaffolder, flag-guarded) is the last.**
+
+## 0.62.0 — 2026-06-20
+
+- **`/red-team` — turn BOSS's defenses into evidence (IDEA-033 #3; the "defense → measured" Anthropic
+  move).** `agent-security` is *prevention* (the deny-list floor, secrets-guard ceiling, Rule of Two);
+  `/red-team` is *proof*. A new **L1 skill** that adversarially tests an AI-mediated FEAT (or BOSS's own
+  conscience hook, `--self`) against the **OWASP 2025 LLM Top 10** — prompt injection (direct +
+  indirect), sensitive-info disclosure, improper output handling, excessive agency, system-prompt
+  leakage, vector weaknesses, misinformation, unbounded consumption, supply chain, data poisoning. Each
+  category gets a **binary pass/fail + the attack that proved it**; **failures become `/evals` cases**
+  (defense → test → regression-proof), pairing with `/evals` (correctness) and the agent-security
+  practice (prevention). Cohort-aware (domain-expert gets the full battery + escalation route;
+  first-product gets the high-value subset in plain language). Honest scope line every run: red-teaming
+  lowers risk, it doesn't certify safety; the deterministic deny-list floor stays the load-bearing
+  prevention. Registered in L1; `boss map` + cheatsheet updated; eval 105/0.
+
+## 0.61.0 — 2026-06-20
+
+- **Scouted skillsmp.com; routed two Anthropic skills through `/vet` → `/boss-learn` (dogfood).** Five
+  marketplace skills reviewed against BOSS's own machinery; the on-principle move was to *vet, not
+  hand-absorb*. Two earned an ADAPT, three did not. **What landed:**
+  - **`library/practices/skill-authoring.md`** (new) — from Anthropic's `skill-creator` ([RVW-013](../docs/research/verdicts/RVW-013-skill-creator-authoring-discipline.md),
+    ADAPT). Fills a real void: BOSS authors skills as its core motion but had *no* written authoring
+    discipline. Captures the three transferable principles — explanatory-over-prescriptive (the
+    IDEA-014 / Principle #2 stance applied to *how we write skills*), progressive disclosure, and
+    descriptions that earn their triggers — plus a ship-time self-check. The heavy with/without
+    **eval-harness is deliberately left out** (duplicates `/vet` + `conscience-evals/`); deferred to IDEA-033.
+  - **`design-system.md` → "Aesthetic ambition — past the slop default"** — from Anthropic's
+    `frontend-design` ([RVW-014](../docs/research/verdicts/RVW-014-frontend-design-aesthetic-ambition.md),
+    ADAPT). The practice owned the *discipline* axis (tokens, the 47 blues, missing states) but was
+    silent on the *taste* axis. Adds the anti-AI-slop stance, the five aesthetic dimensions, and a
+    one-paragraph design-thinking pre-pass — **bounded by BOSS's restraint** (a11y + five states + perf
+    are floors; minimalism is the safer default for a green founder, against the source's maximalist lean).
+  - **Three rejected, recorded:** `ui-ux-pro-max` (checklist mined into IDEA-033; CLI+DB machinery
+    rejected — zero-dep ethos), obra/superpowers `brainstorming` (its "every project, no exceptions"
+    is the literal anti-thesis of BOSS's JIT bet; two micro-techniques harvested to IDEA-033),
+    `code-reviewer` (already dominated by Claude Code's own `/code-review`).
+  - IDEA-033 backlog extended (items 6–8: skill-eval harness, UI/UX pre-delivery checklist, `/spec`+`/consult`
+    question-discipline audit) — each earn-it-gated, not green-lit.
+
+## 0.60.0 — 2026-06-20
+
+- **`docs/PATTERNS.md` — the patterns writeup (the packaging "cool" move, documented).** A public-facing,
+  builder-audience doc that names the engineering patterns BOSS is built on, **framed in Anthropic's own
+  2026 vocabulary with real numbers** — the highest-resonance, lowest-effort move from the Anthropic-appeal
+  research (it's *packaging* what BOSS already has, not new engineering):
+  - **The conscience separates the doer from the judge** (their #1 2026 motif) — a deterministic
+    `UserPromptSubmit` hook (438 lines, **zero model calls of its own**) that gates, then hands the model
+    a *bounded* read in a *fresh* context. Unprompted + isolated.
+  - **Two eval surfaces, with numbers:** 105 deterministic gate cases / 0 failures + **24 GRADED**
+    LLM-judge cases (separate pass, transcripts read).
+  - **Progressive-disclosure skills:** 29 skills, **~1.7k tokens avg** (< Anthropic's 5k guidance),
+    loading JIT (~100-token description until invoked).
+  - Dormant-by-default hooks + frequency-not-tokens cost ledger; security lineage (deny floor →
+    secrets-guard ceiling → Rule-of-Two) stated honestly; AGENTS.md portability; persona-as-both.
+  - **The honest limits are not buried:** zero real founders (its own canvas's 100%-risk), the conscience
+    is Claude-bound, a synthetic judge is still synthetic. (Anthropic rewards honest framing over theater.)
+  Linked from the README ("building agent tooling yourself?"). Ships (tracked doc). No behavior change.
+
+## 0.59.0 — 2026-06-20
+
+- **Venture-brain living memory — the write/evict side (IDEA-022 Track 0; the research's #1 capability
+  gap).** The brain spine (`boss brain` + `record`, model-owned `read.md` + CLI-owned `index.json`)
+  had the *read* side; this adds the **write/evict** side the 2026 gaps research named as the biggest
+  upgrade (Anthropic's memory tool + context editing: +39% / 84% fewer tokens) — and which `brain.js`
+  itself named as next. **Living memory ≠ infinite memory:**
+  - **`boss brain --diff`** — the read's *evolution* (date + headline per session, from the index):
+    continuity made visible without dumping the whole prose.
+  - **`boss brain forget --before <date>`** (or `--id <bN>`) — the **evict** side: drops dated reads
+    older than the date from `read.md` + prunes matching index entries, **keeping the standing summary
+    (preamble) + recent reads**. Founder-invoked, *never automatic* — it's an opinion about a person,
+    so only the human prunes it (on-principle).
+  - **Recency-window gate** (8 sessions) — `boss brain` nudges toward compression/eviction when the
+    read gets long, so the always-loaded surface stays lean (the bloat the conscience itself warns
+    against).
+  - **`/close` pairs the model side:** when the read spans many sessions, fold the oldest reads'
+    lasting conclusions into the standing summary at the top, drop the verbatim old blocks. The
+    standing summary survives; dated blocks are working history that ages out.
+  Additive to the parallel-session spine (no existing behavior changed); zero-dep, format-based block
+  handling (CLI owns boundaries, model owns content); verified in `/tmp` (diff, evict-preserving-
+  preamble, gate); eval 105/0. **IDEA-022: Tracks 1+2 + the Track-0 spine's living-memory increment now
+  shipped; Track 4 (fuller voicing) + Track 3 (AI-native scaffolder) remain.**
+
+## 0.58.0 — 2026-06-20
+
+- **Scaffold `AGENTS.md` — host-neutral, fixes a self-contradiction (IDEA-032).** The clearest cheapest
+  miss from the 2026 gaps research: BOSS scaffolded **Claude-only `CLAUDE.md`**, locking every venture
+  to one host — against its own [IDEA-006](../docs/ideas/IDEA-006-conscience-host-portability.md)
+  host-portability principle + Principle #5 (optionality). Fixed the way **Anthropic's own docs
+  recommend** (verified via `code.claude.com/docs/en/memory`): `AGENTS.md` carries the host-neutral
+  working rules + conventions (read directly by Codex/Cursor/Copilot/Devin/…); `CLAUDE.md` is a thin
+  Claude-specific layer that **imports it via `@AGENTS.md`** (loads into context at launch — no
+  duplication, no drift) and adds the skills/conscience/mode-ladder below. `boss new` emits both;
+  **`boss adopt` handles every case non-destructively** — bare repo → both copied; existing CLAUDE.md
+  → preserved + an adopt block that `@AGENTS.md`-imports the (now-present) rules; existing AGENTS.md →
+  preserved + BOSS's working rules appended as a marked block. New reusable `appendMarkedBlock` helper.
+  Verified all four cases in `/tmp`; eval 105/0; `npm pack` ships `AGENTS.md`. **A concrete
+  down-payment on IDEA-006 + an Anthropic-appeal signal (portability is the property they evangelize).**
+
+## 0.57.0 — 2026-06-20
+
+- **`/consult` — convene the mentor board on a cross-cutting question (IDEA-022 Track 2).** Some
+  decisions don't belong to one mentor (raise-vs-bootstrap touches fundraising + business + GTM +
+  humane at once). A new **L1 skill** that orchestrates the individual `mentor-*` agents: route the
+  question to **only the mentors with a stake** (read the installed board from `.boss/manifest.json` —
+  it grows by mode), get each take **in its own lens** (grounded in the canvas/RESUME/FEAT, pushback
+  included), then **synthesize with the disagreement kept visible** — *where seasoned advisors disagree
+  is where the real decision lives; never average them into mush.* `mentor-humane` keeps its standing
+  **override authority**; the synthesis ties back to the canvas's riskiest assumption and **hands the
+  call back to the founder** (advisory, never a gate; record which lens you followed + why). Reads/
+  writes the venture brain (`.boss/brain/`) when present (IDEA-022). Registered in L1; `boss map` +
+  cheatsheet updated; eval 105/0. **IDEA-022 progress:** Track 1 (`boss adopt`, v0.56) + Track 2
+  (`/consult`, v0.57) now shipped; Tracks 3-4 (AI-native scaffolder, venture-brain voicing) remain.
+
+## 0.56.0 — 2026-06-20
+
+- **`boss adopt` — bring BOSS into an already-started repo, non-destructively (IDEA-005; Track 1 of
+  IDEA-022).** The largest realistic adoption path — most founders have a repo *before* they hear about
+  BOSS — and the one people kept asking for. `boss adopt [--mode <m>]` in an existing dir: copies only
+  what doesn't collide (a new `cpSafe`/`applyStageSafe` — **never `cpSync`'s clobber**), merges the
+  conscience-hook registration into an existing `settings.json` *additively* (the founder's permissions
+  + their own hooks preserved), appends a small marked block to an existing `CLAUDE.md` (or copies the
+  template's if there is none), stamps `.boss/` as **not-self-hosted + adopted**, and registers it.
+  **"Lite BOSS" is the design, not a fallback (Principle 2):** defaults to the lightest register
+  (Quickstart); `--mode mvp` adopts the *full chain* (lays down Quickstart's foundation too, exactly as
+  `boss new` + `boss unlock mvp` would) for a brownfield app that's already earned it. Idempotent
+  (refuses re-adopt → points at `boss sync` / `boss unlock`). Once adopted it's a normal registered
+  project on the usual sync loop — no special-casing. Verified end-to-end in `/tmp`: a real repo's
+  `CLAUDE.md`, `settings.json` (custom permission + `Stop` hook), `README`, and `src/` all preserved;
+  conscience loops + hook land so it can fire; `boss map` works; `--mode mvp` lays L0+L1. eval 105/0.
+  Wired into `boss --help`. **This is also IDEA-022 Track 1** — the living conscience needs a venture
+  to read, and adopt is how an existing one gets a venture brain.
+
+## 0.55.0 — 2026-06-20
+
+- **`/persona` — your app's target-user as a consultable agent voice (IDEA-031).** Occasioned by
+  Ajesh: *"if I wanna build an app for moms to track chores, the first persona is moms… we can do a
+  Q&A with the builder, online research, or a UX researcher drops their research in… the app uses it
+  as an agent voice to guide product decisions"* — *"its also the QA, its both right?"* A new **L0
+  skill**: **derive** the primary target-user persona from the idea → `docs/personas/<slug>.md`
+  (who · context · jobs · pains · values · *what we don't know yet* · synthetic/real evidence ledger);
+  **enrich** from four sources (builder Q&A · `deep-research` online · drop-in real research via
+  `/import` · passive read of idea/canvas); **consult** in voice in **both directions** — *guidance*
+  ("would she want X?") and *QA* (Husain-discipline structured reactions on a build, comparable across
+  versions). **The discipline is the product:** every consult is framed as a *pre-filter, never
+  validation* — balances interest with concerns, names its blind spots, and closes with the
+  go-ask-a-real-one caveat (Fitzpatrick/Mom Test); synthetic shrinks as real grows, visibly. Reuses
+  [IDEA-009](../docs/ideas/IDEA-009-proto-personas-as-evolving-instruments.md)'s evolving-instrument
+  methodology pointed at the *founder's* users (vs. BOSS's internal cohort instruments). Registered in
+  L0; `boss map` + cheatsheet updated; eval 105/0.
+- **`/prototype` refined by its own persona-reactions pass (IDEA-009 instrument working).** Ran the 4
+  BOSS-internal personas on the new features (`/prototype`, kanban, upstream conscience); the
+  convergences drove real fixes: (1) the after-run nudge now leads with a **concrete plain-language
+  action** and lets *"I don't know yet"* be a fine answer (beginners bounced on `/canvas` /
+  "pressure-test" jargon); (2) the **5-token pass is skipped on throwaway sketches** by default
+  (eng-builder caught it contradicting "tangible beats pretty"); (3) **`--stack=<x>`** is now a
+  first-class option + the stack pick is narrated in plain words for beginners; (4) trimmed the
+  defensive over-explanation (dropped the "not vibe coding" protest; sketch-vs-MVP named once; the
+  "becomes the MVP" rule now names the *real* failure — bolting auth onto throwaway code). The
+  build-integrated eval channel caught design issues before a real founder hit them — first evidence
+  for IDEA-009's claim #1.
+
+## 0.54.0 — 2026-06-20
+
+- **Upstream conscience — `/spec` now asks "is it worth building?" not just "is it built right?"
+  (IDEA-026 Part A, closes IDEA-026).** The biggest *conceptual* delta from the 2026 leader scan
+  (Ng's "PM bottleneck," Appleton's "align before the agent runs"), shipped the small + safe way the
+  host-subtraction audit found: **not a new always-on loop, a voice-sharpening of the existing
+  `spec-loop` restraint** (which already fires, skill-invoked, when the founder reaches for `/spec`
+  before the canvas closes). The restraint frame now surfaces the *substantive* gap — *who is this
+  for, and what's the bet that could sink it?* — not a checklist. **Respects `/prototype`:** a
+  throwaway sketch needs none of this (build-first is legitimate); the question fires only when
+  committing to build *for real*. eval gate 105/0 (skill-voice change, no predicate change).
+- **Positioning reframe — README opener now leads with the judgment gap (option C).** *"Everyone can
+  build now. Almost no one can tell a real business from a convincing demo. BOSS is the conscience
+  that keeps you honest while you move fast."* + the floor/ceiling subline (*vibe coding gets you a
+  demo; the discipline on top gets you a business*). The 2026-vocabulary update to positioning
+  (harness-is-the-moat / agentic-engineering-is-the-discipline), in BOSS's voice. Full draft +
+  rationale in the gitignored `docs/dossier/positioning-reframe-2026.md`.
+
+## 0.53.0 — 2026-06-20
+
+- **`/judge-traces` — the deliberate reader for the trace substrate (IDEA-025 Phase 2).** v0.48 shipped
+  `auto-log` (collects `.boss/trace.jsonl`); this is the skill that *reads* it — completing a
+  shipped-but-inert capability (a substrate nothing reads has no purpose). A new **L1 skill** applying
+  Hamel/Shankar's 2026 discipline to the founder's *own* sessions: read real traces → factual shape
+  first (cheap, deterministic) → sort failures into a **binary** pass/fail taxonomy (`wrong-files`,
+  `thrash`, `silent-scope-creep`, `no-trace-of-the-point`, or your own) → route *recurring* modes to
+  `/boss-learn`. **One expert not a committee; don't grade its own homework (judge the trajectory);
+  counts are the signal.** Graceful degrade: no trace → honest "nothing to judge yet, turn on
+  `auto-log`," never a fabricated taxonomy. **Collection ≠ judgment** stays load-bearing: `auto-log`
+  collects passively, `/judge-traces` judges deliberately — never fused into an always-on auto-grader.
+  Registered in L1; `boss map` + cheatsheet updated; eval 105/0.
+- **Host-subtraction audit drafted (IDEA-028) + an IDEA-026 Part A finding (workspace docs).** The
+  `mentor-architect` pass concluded **retire nothing**: BOSS's conscience fires *unprompted/event-driven*,
+  and native `/goal`/`/loop` are *user-invoked* — they can't replace a conscience that speaks when you
+  wouldn't have asked, so the host did *not* absorb the loop runtime (the moat). Sit-on the host for
+  future orchestration (dynamic Workflows) + the secrets ceiling (auto-mode hard-deny); keep the
+  `permissions.deny` floor (more portable). **Bonus finding:** `spec-loop` already implements most of
+  the "upstream conscience" (IDEA-026 Part A) — it fires restraint when the founder reaches for `/spec`
+  without the canvas closed — so Part A is a *voice sharpening of spec-loop*, not a new `worth-building-loop`
+  (which would redundantly overlap spec-loop + caution + drift). Left for Ajesh's eye on the exact
+  voice (it touches conscience tone). No code retired; decision-input only.
+
+## 0.52.0 — 2026-06-20
+
+- **`/prototype` — drop an idea, hit go, see something tangible (IDEA-030).** Ajesh's framing settled
+  the design: *"not just vibe coding, but not gatekeeping until so much thought… building first in a
+  lean cycle is a place to start, not waiting until the other two are clear. People can fill in the
+  missing pieces after they get the gist out of their head and see something tangible."* A new **L0
+  skill** that builds the smallest runnable, clickable version of an idea — the ONE core interaction,
+  in whatever stack gets to "click it" fastest, mock data freely, the 5-token distinctiveness pass so
+  it doesn't look generic — then runs it. **The load-bearing call:** the conscience fires **AFTER** the
+  thing runs ("there it is — does seeing it change the idea? when you're ready: `/canvas`"), **never a
+  gate before** — building first *is* a legitimate first move in the loop, not a skip of it. Cohort-
+  aware (first-product gets the magic moment; eng-builder gets stack control; domain-expert gets the
+  "this is a sketch, not a regulated tool" guardrail up front). Honest framing held: it's a *sketch to
+  think with, named once, not your MVP* — and the graduate ladder (`unlock mvp` → `/spec` → `/evals`)
+  is what keeps a fast prototype from becoming a pseudo-app (PRINCIPLES). Registered in the L0
+  manifest; `boss map` + cheatsheet updated; eval 105/0.
+
+## 0.51.0 — 2026-06-20
+
+- **Visual kanban (`boss board --html`) + a voice-tightening pass.** Two founder-experience asks.
+  - **`boss board --html` — the board, as a visual kanban.** Ajesh: *"having a visual kanban state
+    that can be updated when the board is is super helpful."* Promotes the HTML view IDEA-015 deferred
+    behind an earn-it gate. Same **pure projection** as the terminal board (`collectBoard`) rendered to
+    a self-contained `.boss/board.html` — zero deps, no server, no JS framework; four columns
+    (Captured / Taking shape / Building / Shipped), cards with id + title, `↻ review due` + `blocked`
+    flags, the evidence line + riskiest-assumption framing on top, light/dark, responsive. Opens in
+    the default browser (best-effort; the printed path is the contract). "Updated when the board is" =
+    re-run it — it's a read of the files, never a maintained doc (the IDEA-015 discipline holds). Wired
+    into `boss --help` + `boss map`.
+  - **Voice pass (voice-keeper full audit — verdict: "in remarkably good shape").** Applied the
+    high-leverage fixes: dropped the `🎯` emoji that shipped into the founder's own canvas template;
+    unified the cohort question wording so `/welcome` and `/boss` are *actually* identical (they
+    claimed to be); settled the self-description on **"build tool"** (not "build companion" — the
+    ethos is conscience/tool, not coach); collapsed the most-read `boss new` first-run lines to one
+    bold path + demoted fallbacks (matching `/welcome`'s own discipline); collapsed the `boss insights`
+    triple-stated telemetry footer; dropped an inside-baseball `/ai-cost` wink in `boss conscience
+    cost`; de-"coach"-ed the README mentor line against the settled ethos; added `boss brain` +
+    `boss insights` to the `boss map` standing-controls so the live cheatsheet agrees with `--help`.
+    (Kept the maintainer-only `Learned … UP` vocabulary deliberately — UP/DOWN is load-bearing
+    Principle-1 language and the reader there is always the maintainer.)
+  - **Captured (not built): [IDEA-030 "drop an idea and hit go"](../docs/ideas/IDEA-030-drop-an-idea-hit-go.md)** —
+    a fast path to a runnable prototype. Philosophically loaded (the pseudo-app trap is *why BOSS
+    exists*), so captured with a concrete `/prototype` proposal + the resolution (see-it-not-sell-it,
+    conscience-attached, the graduate ladder as the honesty mechanism) and four shape questions for
+    Ajesh — a decision to take together before building.
+  - Eval gate 105/0; judgment GRADED; HTML render + voice fixes verified end-to-end in `/tmp`.
+
+## 0.50.0 — 2026-06-20
+
+- **Close the trends-pass loose ends (IDEA-027 #4 + IDEA-026 Part B wiring).** Ajesh's audit prompt —
+  *"did we miss anything else to implement? you did a lot of lookup."* Two genuine loose ends from the
+  research, both now closed:
+  - **`boss board` surfaces "review due" — the trigger half of `/revalidate`.** v0.48 shipped the
+    revalidation *gate* but nothing *surfaced what to revalidate* — a half-feature. The board now reads
+    each item's `next_review:` frontmatter and flags any whose date has passed (`· ↻ review due`) plus
+    a footer line `↻ N past review — run /revalidate <id>`. **Frontmatter-true, never guessed:** no
+    `next_review` date → not flagged (an age-inferred "stale" signal would be noise the founder learns
+    to ignore). Completes the dhun-ported revalidation lifecycle; `boss board` stays a pure projection
+    (no new state).
+  - **Agent security wired JIT into `/ai-first-init` (Step 5.5).** v0.49 shipped `agent-security.md` as
+    an UP practice, but founders don't read `library/` — so the lethal-trifecta / Rule-of-Two framing
+    now surfaces *in the AI-native day-one sequence* (one sentence on a Quickstart; more ceremony as
+    the app reads untrusted input / handles regulated data). Names the surface, points at the
+    `permissions.deny` floor + `secrets-guard` ceiling + the deterministic-guard rule.
+  - Eval gate 105/0; judgment GRADED + no blocking failures; board staleness verified end-to-end in
+    `/tmp` (past-review flagged, future-review not). **Honest remaining map (deliberately staged, not
+    missed):** IDEA-025 P2/P3 (`/judge-traces` + trace-fed learn loop — need accumulated traces),
+    IDEA-026 Part A (upstream conscience — needs the IDEA-028 host-mount), IDEA-028 host-subtraction
+    audit (a decision to take *with* Ajesh, not autonomously), AGENT_DOC_MAP (until two agents
+    collide), the SDD-vocabulary + "agentic-engineering" positioning reframes (mentor-pitch territory),
+    and publishing the library on the open Skills standard (IDEA-006 distribution).
+
+## 0.49.0 — 2026-06-20
+
+- **Embed the 2026 best-practices — design, evals, security feel current across the board (IDEA-029 +
+  IDEA-026 + the trends pass).** Ajesh's follow-on to v0.48: *"on all research you surfaced (not just
+  design), let's ID all critical to add to our current work, and just go and embed it… make BOSS feel
+  very up to date with best practices."* Where v0.48 shipped the dhun *machinery*, this embeds the
+  *thinking* into the surfaces founders actually touch. Provenance:
+  [SESSION-2026-06-20-ui-design-scan](../docs/research/sessions/SESSION-2026-06-20-ui-design-scan.md).
+  - **`/design-tokens-init` — the 5-token distinctiveness pass + DTCG/semantic naming (the design
+    win).** Three layers already prevented *drift*; the new section prevents *sameness* — the
+    "shadcn trap" (slate/Inter/8px/indigo = generic-AI-app look) broken by five deliberate overrides
+    (warm neutral · intentional radius · type pairing · one owned accent · **one "signature token"**),
+    cohort-aware. Plus: name semantic tokens by **purpose not hue**, emit **W3C DTCG** where the stack
+    allows (portability, no lock-in), and **inline a semantic→primitive map into CLAUDE.md** so the
+    agent inherits the brand on every turn. (freedesignmd · Vercel · Curtis · W3C DTCG.)
+  - **`/evals` — the 2026 Hamel/Shankar sharpening.** Error analysis on **real traces** first (and a
+    pointer to v0.48's `.boss/trace.jsonl` as that raw material); **binary pass/fail, not 1–5 scores**;
+    **one expert, not a committee**; **don't let the model grade its own homework** (separate verifier
+    + trajectory-not-just-endpoint); a **60/30/10** deterministic/judge/human cost hierarchy.
+  - **New UP practice `library/practices/ai-ux-patterns.md`** — the interaction layer (where
+    `design-system.md` owns the look): "why this" rationale grounded in the user's own inputs ·
+    confidence-as-register · **Notify/Question/Review** interrupt registers · **edit-before-execute +
+    risk-tiered gates** (four decision verbs; gate by loss type) · **trust-repair after a miss**
+    (asymmetric recovery) · progressive disclosure · **discernment — knowing when not to speak — as a
+    first-class fundamental** · pinned canonical refs (Shape of AI / HAX / IBM Carbon; community
+    catalogs via `/vet`). Captured as [IDEA-029](../docs/ideas/IDEA-029-ai-native-interface-patterns.md),
+    extends IDEA-010.
+  - **`docs/mentor-practitioners.md`** AI-UX heuristics block updated with the 2026 additions +
+    pointer to the new practice.
+  - Library-layer + skill-content changes (no new template skill, no CLI change); eval gate 105/0,
+    judgment GRADED + no blocking failures; `boss new` + `unlock mvp` verified the upgraded skills
+    ship. **Staged from the design scan:** planning-as-collaboration / mid-run steering (Appleton —
+    twin of IDEA-026 Part A), RESUME-as-agent-inbox, wrapper-vs-flatten-per-cohort.
+
+## 0.48.0 — 2026-06-20
+
+- **The 2026-trends + dhun-scan pass — first builds (IDEA-025/027/026).** Occasioned by Ajesh:
+  *"with all the latest trends since we built BOSS… as well as potential new methods inside dhun, let's
+  assess what BOSS can better improve upon."* Three research passes (external 2026 trends; a 2026-only
+  scan of the named practitioners in `docs/mentor-practitioners.md`; a full catalog of the sibling
+  **dhun** project's working-method machinery) → captured as
+  [IDEA-025](../docs/ideas/IDEA-025-trace-native-conscience-and-evals.md)…028 with full provenance in
+  [SESSION-2026-06-20](../docs/research/sessions/SESSION-2026-06-20-trends-and-dhun-scan.md). This
+  release ships the concrete, proven, low-risk slice; the conceptual reframes (upstream conscience,
+  host-subtraction) stay specced-and-staged. **Three new dormant/UP capabilities — no behavior change
+  until opted in:**
+  - **`auto-log` trace substrate — the keystone, shipped dormant (IDEA-025 Phase 1).** A zero-dep Node
+    SubagentStop hook (`library/hooks/auto-log.js` + L1 template) that appends one honest line to
+    `.boss/trace.jsonl` per writer-subagent — session, agent, files actually changed (reads
+    `git status --porcelain`, so it catches *new* files too, not just tracked diffs), with last-line
+    dedup and read-only-agent skip. **The within-session complement to v0.47's `boss insights`**
+    (cross-project registry): both honest-trace, local-only, append-only, measure-don't-instrument
+    (inherits the IDEA-021/013 contract). It's the raw material a future trace-native judge
+    (`/judge-traces`) + sleep-time learn loop read — Hamel ("error analysis on real traces") + Chase
+    ("traces, not code, are the source of truth"). **Dormant** (a SubagentStop hook costs per-subagent
+    latency — registration is the opt-in, same as `secrets-guard`).
+  - **`memory-cue` hook — feedback→memory nudge, shipped dormant (IDEA-027 #1).** Ported UP from the
+    dhun dogfood, Node-ported for the zero-dep rule (`library/hooks/memory-cue.js` + L0 template). A
+    UserPromptSubmit hook that regex-detects a feedback signal (directive / corrective / confirmation)
+    and *nudges* the model to save it to memory — never auto-writes (wording needs reasoning), silent
+    on no-match (zero token cost). Serves the `library/memory-seed/` ambition.
+  - **`/revalidate` — the 3-line gate against zombie features (IDEA-027 #2).** A new L1 skill (+
+    `library/practices/revalidation.md`) ported UP from dhun's REVALIDATION lifecycle: before paused
+    work re-enters the build, answer *still relevant? still aligned? anything changed?* → revive /
+    rescope / kill / re-pause. BOSS eats it first — `docs/RESUME.md` carries a long deferred list.
+  - **Two new UP practices** distilled from the same scan: `library/practices/quality-ratchet.md`
+    (dhun's `.ratchet` one-way-baseline pattern, stack-neutral) and `library/practices/agent-security.md`
+    (Simon Willison's 2026 lethal-trifecta / "Agents Rule of Two" / "deterministic guard around a
+    non-deterministic model" — IDEA-026 Part B, the next ring after `secrets-guard`).
+  - Zero-dep held (Node built-ins only; `npm pack` verified — all new files ship). Hooks dormant
+    (settings.json unchanged; only `conscience.js` registered). `boss map` shows `/revalidate`;
+    eval gate clean (no blocking failures); judgment GRADED 7/7; tested end-to-end in `/tmp`
+    (`boss new` + `boss unlock mvp`, both hooks present + dormant, skill present).
+
+## 0.47.0 — 2026-06-19
+
+- **Humane two-way learning channel — built the moment BOSS went public, the humane way (IDEA-024).**
+  Going public (MIT, github.com/ajeshh/bossbuild) turned a private dogfood into a thing strangers
+  run, which needs a way to learn + pivot. Ajesh asked for "feedback from end-users, and learn
+  *passively* how users use it." The second half is the exact surveillance line BOSS exists not to
+  cross ([IDEA-021](../docs/ideas/IDEA-021-passive-instrumentation-and-fleet-learning.md)). Applied
+  BOSS's own conscience (mentor-humane fork; PRINCIPLE: humane before viable) and built the honest-trace
+  version instead — **no silent telemetry.**
+  - **`/feedback` (direct, user-initiated).** A founder-facing skill in L0: send a bug / confusion /
+    wish back upstream. Shows the founder the exact title + body + the *one* line of context
+    (`BOSS <ver> · <mode> · <OS>`) before anything leaves the machine; files a GitHub issue via
+    `gh issue create`, or falls back to a prefilled issue link to paste. Public-repo warning stated.
+    Never automatic, never a hook.
+  - **`boss insights` (passive, the humane way).** Reads the trace your own work *already leaves* —
+    your registered projects on *this machine* — and reports where each venture's loop stands
+    (idea → canvas → build), flagging empty / untested / stale. **Measures graduation + loop-closure,
+    never activity/engagement** (the vanity metric BOSS refuses to expose). Local-only; nothing is
+    sent. Zero-dep (`src/insights.js`).
+  - **Opt-in share-up contract.** New `shareUp: false` default in `.boss/config.json` — any future
+    cross-user learning is gated on this flag being true *and* a per-send confirmation, by construction.
+    Telemetry is never a default.
+  - **Not built (deferred, named in IDEA-024):** silent cross-user telemetry (the line — not crossing
+    it); a full `npm publish` / auto-update pipeline (premature at n≈1 — `boss status` already nudges
+    on version drift; the real risk is demand, not distribution). Captured in
+    [SESSION-2026-06-19-founder-test](../docs/research/sessions/SESSION-2026-06-19-founder-test.md).
+
+## 0.46.0 — 2026-06-19
+
+- **Bring-your-own-material import — the on-ramp from "I jotted it somewhere" (IDEA-023).**
+  Occasioned by a live founder-test: a founder ran `boss new`, then stalled — the idea lived in a
+  Word doc / Google Doc / Obsidian note / PDF / deck / URL, and there was **no way to bring it in**.
+  A correctly-scaffolded but idea-less project read as *"empty… I'm stuck."* This closes that
+  first-run dead-end.
+  - **Load-bearing decision:** import lives in the **skill layer, not the zero-dep CLI** — the CLI
+    (Node built-ins only, Principle 4) can't parse PDF/docx or fetch a URL, but the model already
+    reads heterogeneous formats natively. Same predicate/runner split as the loop runtime (IDEA-008):
+    deterministic core stays deterministic; the model does the parsing + shaping.
+  - **What ships:** (1) `/boss` §1 now ingests **one-or-more sources** — local files *and* URLs, in
+    any mix — snapshots a durable copy of each into `docs/source/` ("the project owns a copy"), and
+    synthesizes across all of them before shaping the idea. (2) A new **`/import`** skill (registered
+    in the L0 manifest) for adding material to an *already-captured* idea, or as an alternate spin-up
+    door. (3) **Discoverability fix** — the part that actually stuck the founder: `boss new`'s "Next:"
+    block now shows `code <name>` (editor handoff) + the file/url/import options; `/welcome` (both the
+    full tour and the 30-second version) and the L0 `CLAUDE.md` template all advertise bring-your-own
+    material.
+  - **Deferred (named in [IDEA-023](../docs/ideas/IDEA-023-bring-your-own-material-import.md)):**
+    material-first ordering (point at material → BOSS names the folder), binary/OCR formats
+    (`.pptx` text, image-only PDFs), live-source re-pull vs. one-time snapshot, and a CLI
+    `boss import` second door (only if the skill path proves it's wanted outside Claude).
+  - Dogfood target: `~/Projects/fraands` (the project that surfaced the gap). Surfaced + captured in
+    [SESSION-2026-06-19-founder-test](../docs/research/sessions/SESSION-2026-06-19-founder-test.md)
+    (OBS-002/003/005).
+- **`/welcome` closes on the action — long content no longer buries the next step (OBS-001).**
+  Same founder-test: *"the welcome message is a bit too long… I forget what I'm supposed to do next."*
+  The skill already had an "end on one next step" rule, but the beginner tour printed three full
+  reference sections (conscience / modes / help) *after* the next step, walling it off. Fix: a new
+  voice rule ("close on the action — long content must tie back to the next step"), and a structural
+  **pivot** — after the shape + next step, the three reference sections are now **offered, not dumped**
+  (tagged `reference — expand only if asked`); the founder leaves on `/boss` or `/triage`, not on a
+  wall. (OBS-004 — host-switching across Claude app/VSCode/Cursor — logged, deferred to IDEA-006.)
+
+## 0.45.0 — 2026-06-05
+
+- **JIT working-context, Phase 1 — every `boss new` project is now JIT-by-construction (FEAT-020).**
+  The deny-list (v0.42) made projects secrets-safe by default; this makes them *context-lean* by
+  default. The principle was already vetted (`context-discipline` practice, RVW-005/010) but only
+  *described* path-scoped rules — now the templates **ship** them. This is Phase 1 of a 4-phase
+  lifecycle ([`docs/ideas/FEAT-020`](../docs/ideas/FEAT-020-jit-working-context-lifecycle.md)); Phases
+  2-4 (`/close` GC, promote-on-evict via `/extract`, a freshness moment) are specced + deferred with
+  triggers.
+  - **What ships:** `.claude/rules/` examples with `paths:` frontmatter that load **only when Claude
+    opens a matching file** (not at session start) — L0 `your-app-code.md` (the basic path-scoped
+    pattern), L1 `feature-context.md` (the live feature's working notes, which Phase 2's `/close` will
+    later compress to a one-liner). The two-memory cut that decides what goes where is documented in
+    the new `library/memory-seed/` shelf (README + an example durable-facts seed). L0 `CLAUDE.md`'s
+    Memory line now names both halves (durable → auto-memory; working-state → path-scoped rules).
+  - **Verified before building:** confirmed against the official Claude Code docs (2026-06-05) that
+    `.claude/rules/` with a `paths:` key is real and JIT-loaded — *not* a confusion with Cursor's
+    `.cursor/rules` (`globs:`). The `context-discipline` practice's "re-verify host syntax on every
+    build" rule, honored.
+  - **The restraint line (carried from the IDEA):** Phases 2-4 risk being BOSS gold-plating its own
+    substrate; the recency-window-by-hand is currently enough. First dogfood is BOSS's own repo — a
+    Phase-1 slice going stale and misinforming a session is the cleanest Phase-2 re-open signal.
+  - Zero-dep held; tested end-to-end in `/tmp` (`boss new` → rule present; `boss unlock mvp` → feature
+    rule added). `mentor-architect` pass + 4 forks decided with Ajesh recorded in the FEAT.
+
+## 0.44.0 — 2026-06-02
+
+- **`secrets-guard` PreToolUse hook — the high-stakes ceiling, shipped opt-in (closes the RVW-005
+  follow-on, the principled way).** The v0.42 deny-list is the universal zero-cost floor; this hook is
+  the broader-coverage ceiling — and the v0.42.1 reconsideration said it must NOT be a universal
+  default (a `PreToolUse` hook spawns a process on *every* tool call). So it ships **dormant**:
+  `library/hooks/secrets-guard.js` (canonical) + `.claude/hooks/secrets-guard.js` in the L0 template,
+  **not registered by default.** Registration is the on-switch (an unregistered hook costs nothing),
+  recommended for the `domain-expert` / regulated cohort.
+  - **Behavior:** Read/Edit/NotebookEdit of a secrets file (`.env`/`.env.*`/`secrets/**`) → **deny**
+    (reading secret contents into context is the leak); Bash or MCP referencing a secrets path →
+    **ask** (don't hard-block legit `.env` *creation* — surface it to the human); else allow.
+    **Fail-open** (any parse/runtime surprise → allow; a guard that breaks the session is worse than
+    one that occasionally misses, and the deny-list floor still hard-blocks the common vectors).
+  - **Tested** (10 cases, piped PreToolUse events): denies Read/Edit of `.env`/`.env.local`/
+    `secrets/`; allows `src/app.js`, `npm test`, and `.environment.ts` (no false positive); asks on
+    `cat .env` + MCP-with-secrets; fail-open on malformed input. Zero-dep Node, output per the Claude
+    Code PreToolUse contract (JSON `permissionDecision`, exit 0).
+  - **Cohort auto-registration deferred** (a clean follow-on): wiring `domain-expert` cohort setup to
+    register it automatically. Today it's documented opt-in (snippet in the file header + the
+    `context-discipline` practice).
+
+## 0.43.0 — 2026-06-02
+
+- **Wayfinding, Pass 1 (IDEA-018) — `boss map` + a doc generator that can't rot.** Occasioned by a
+  docs-health pass that found the README **19 releases stale** and, worse, that there was **no "how to
+  use BOSS" guide at all**. The fix is shaped like BOSS itself: wayfinding, not a manual. Decisions
+  locked with Ajesh — **mode ladder is the spine** (persona = entry filter, aspect = which mentor to
+  ask, never chapters); **split audience** (a command ships to founders, the prose stays in the BOSS
+  repo); **durable core first**.
+  - **`boss map`** (CLI, ships to every project) — the *live* cheatsheet. A pure render of state the
+    project already holds (the `.boss` stamp + installed `SKILL.md` files), in the `boss board`
+    spirit: *You are here · available now (grouped by the rung that unlocked each skill) · one unlock
+    away (the next rung's skills, read from the package, with the real project name substituted in) ·
+    standing controls.* Nothing to maintain, nothing to drift.
+  - **The de-rot mechanism** — `src/modes.js` is the single source both `boss map` and the generator
+    read (manifests + `SKILL.md` frontmatter), so the live map and the static docs can never disagree.
+    **`scripts/gen-docs.js`** (`npm run gen:docs`) emits **`docs/CHEATSHEET.md`** (the whole ladder,
+    the wall-poster) and **`docs/SKILLS.md`** (one line per skill, grouped by mode) — both carry a
+    GENERATED banner and are derived, never hand-typed. This is the actual fix for what bit the README:
+    the per-mode lists become a build artifact, not a memory test.
+  - Zero-dep held — `src/map.js` + `src/modes.js` ship; `scripts/` and the generated `docs/` stay
+    dev-only (verified via `npm pack`). Eval suite regression-clean (no blocking failures, GRADED 7).
+    End-to-end tested in `/tmp` (map at Quickstart → unlock mvp → map regroups + previews V1; placeholder
+    substitution + width-capping confirmed). **Pass 2 (the hand-authored `docs/GUIDE.md` walkthrough)
+    is the next session** — written against these generated surfaces, per the agreed sequence.
+
+## 0.42.1 — 2026-06-02
+
+- **BOSS eats its own context-discipline dogfood + sharpens the practice (the learning loop in real
+  time).** Minutes after shipping `context-discipline` (v0.42.0), applied it to BOSS itself and
+  refined it with what doing so taught.
+  - **Dogfood (DOWN):** the recency-window rule (RVW-002) applied to BOSS's own `docs/RESUME.md` — the
+    "State" section was a 25-entry append log read at every session start. Trimmed to the **5 most
+    recent** entries; older versions point to `registry/CHANGELOG.md` (which carries all 43 versions —
+    confirmed before cutting, so it's non-destructive). RESUME dropped **727 → 346 lines.** BOSS now
+    practices the leanness it prescribes.
+  - **Practice sharpening (the part that matters):** reconsidered the deferred PreToolUse secrets-guard
+    hook and recorded *why* it's deferred, not just *that* it is. A `PreToolUse` hook fires a process
+    on **every tool call** (real latency), where `permissions.deny` is a zero-cost native check. So the
+    practice now states: the **deny-list is the universal floor** (always ship); a **secrets-guard hook
+    is a high-stakes/opt-in ceiling** (regulated/PHI cohorts), **not** a universal default — adding
+    always-on per-call machinery for marginal coverage is the framework-bloat BOSS warns founders
+    against (R&H #1 / IDEA-013 cost discipline). This is `/vet`'s skepticism turned inward: even an
+    ADOPT's "ceiling" gets cost-weighed before it ships to everyone.
+  - No template/CLI behavior change beyond the doc + RESUME; the v0.42.0 deny-default still stands as
+    the shipped safe-default.
+
+## 0.42.0 — 2026-06-02
+
+- **`/boss-learn` routes the sweep's first ADOPT — a "context discipline" practice, UP + a DOWN
+  safe-default.** Acting on the v0.41 `/vet` sweep: the two ADOPTs (RVW-005 deny-secrets, RVW-010
+  token-optimization) plus RVW-002 (lean session docs) collapsed into **one** pattern, routed two ways
+  per PRINCIPLE #1.
+  - **Verify-before-encode (the gate both verdicts set).** Before promoting, the version-bound Claude
+    Code claims were checked against current behavior. One was **FALSE — `.claudeignore` does not
+    exist** (the source post conflated it with `permissions.deny`); it was struck from the practice
+    rather than shipped to every project. Confirmed: `permissions.deny` glob syntax (and that a
+    `Read(...)` deny does **not** cover Bash — needs a separate `Bash(...)` rule), PreToolUse hard-block,
+    `.claude/rules/` `paths:` frontmatter, CLAUDE.md load behavior. The `/vet` thesis applied to BOSS
+    itself: popularity ≠ correctness, even when *BOSS* is the one adopting.
+  - **UP** → `library/practices/context-discipline.md`: lean always-loaded docs (CLAUDE.md +
+    RESUME recency-window), path-scoped `.claude/rules/`, `permissions.deny` for secrets *and* bloat,
+    PreToolUse/PostToolUse hooks as the enforcement ceiling. Host-tagged `claude-code` with an explicit
+    "re-verify syntax on host change" note (IDEA-014 recalibration). Provenance cites the RVWs — vetted,
+    not adopted on stars.
+  - **DOWN (product safe-default)** → the L0 Quickstart template now ships a `permissions.deny` block
+    for `.env`/`.env.*`/`secrets/**` (Read + Bash) in `.claude/settings.json`, and `.gitignore` covers
+    `.env.*` + `secrets/`. Every new `boss new` project is secrets-safe by default — the
+    enforce-in-harness principle (RVW-012) made concrete, not left as advice.
+  - **Deferred follow-ons (named, not crammed in — PRINCIPLE #2 / small steps):** a `library/hooks/`
+    PreToolUse secrets-guard (catches Bash + MCP + future skills — code+test, its own step);
+    mode/cohort-scoped `.claude/rules/` in the template; BOSS's own root CLAUDE.md/RESUME trim
+    (RVW-002 — awaiting Ajesh's recency-window size). ADAPTs RVW-007/008 remain founder-facing/scope-gated.
+
+## 0.41.0 — 2026-06-02
+
+- **First `/vet --all` sweep — 10 verdicts (RVW-003…012), and the skill earned its keep.** Ajesh
+  dropped a 10-item pile of AI/Claude-Code "best practices" (Reddit threads + Lenny's-newsletter posts)
+  and swept them in one pass. The distribution is the proof the skill routes on merits, not reflex:
+  **2 ADOPT · 2 ADAPT · 3 NOT-YET · 3 REJECT.** Only 2 clean adopts out of 10 — a skeptic, not a
+  bookmark folder.
+  - **The two ADOPTs collapse into ONE action — a "BOSS context discipline" practice** (the value of
+    synthesizing a sweep instead of N independent verdicts). RVW-005 (hard-deny `.env`/secrets, don't
+    trust prompting), RVW-010 (lean CLAUDE.md <500 tok, path-scoped `.claude/rules/`, `permissions.deny`
+    for bloat, hook noise-filtering), and the earlier RVW-002 (RESUME recency-window) are facets of one
+    practice; RVW-009 (context-engineering failure modes) is its research rationale and RVW-012 (Agent
+    = Model + Harness, "safety lives in the harness not the model") its backing principle. Queued for
+    `/boss-learn` (UP `library/practices/` + `library/hooks/` secrets-guard + DOWN BOSS's own doc trim
+    + template defaults), **gated on verifying version-bound Claude Code specifics first** (IDEA-014
+    recalibration territory).
+  - **Two ADAPTs, both founder-facing + scope-gated:** RVW-007 (Couch-to-5K — adopt the
+    smallest-next-step *philosophy* for `/welcome` + beginner-cohort nudges; **reject the daily-streak
+    mechanic** as the gamification/pressure trap the canvas already refuses — a guardrail recorded so a
+    future session isn't tempted); RVW-008 (categorize-agents / start-simplest — a modest
+    `mentor-architect` framing, strip the enterprise + stack taxonomy per PRINCIPLE #4).
+  - **Three NOT-YET:** RVW-003 (plumbing-awareness — strong founder-facing candidate, re-open when the
+    founder-facing build lands), RVW-009 + RVW-012 (reference pieces — re-open at conscience
+    context-injection review / IDEA-006 host-contract work respectively).
+  - **Three REJECT, recorded with reasons:** RVW-004 (`/remote-control` — out of scope/low-evidence,
+    but kept a humane stance on always-on agent work), RVW-006 (21-hacks listicle — wrong altitude),
+    RVW-011 (n8n tutorial — PRINCIPLE #4, folded into RVW-008 as an example).
+  - **A real skeptical catch:** 6 of the 10 drops were by the **same author** (one newsletter
+    corpus). The sweep applied an **author-concentration discount** — cross-confirmation *within* one
+    voice isn't independent evidence; the "respected practitioner" rubric rung counts once, and the
+    real evidence is the *distinct* sources (the deny-secrets PSA, slaorta, StokeJar, and the
+    DeepMind/Microsoft/Salesforce research cited *inside* the pieces).
+  - Records: `docs/research/verdicts/RVW-003…012`; all 10 inbox items marked `resolved:`. Zero-dep held
+    (`npm pack` ships 0 — all under `docs/`). No code/skill change → gate + judgment suites unaffected.
+    **Nothing is built yet** — ADOPT/ADAPT hand-offs await Ajesh's go (the skill decides *whether*;
+    `/boss-learn` decides *where*).
+
+## 0.40.1 — 2026-06-02
+
+- **`/vet` gains batch sweep — drop a pile, vet once.** From dogfooding `/vet` on real drops (RVW-001,
+  RVW-002): the natural rhythm is *accumulate, then sweep*, not vet-on-arrival. `/vet --all` now vets
+  every un-vetted inbox item — **each as its own full skeptical pass with its own `RVW-NNN` verdict**
+  — then prints one summary table + the ADOPT/ADAPT hand-off list. No-arg `/vet` lists un-vetted items
+  oldest-first and offers the sweep. Already-vetted items (a `resolved:` line or an existing verdict)
+  are skipped. Clarified the old "one claim per run" rule → **"one claim per verdict"**: it always
+  protected *depth-per-claim* (never collapse several claims into one shallow verdict), never forbade
+  vetting many in sequence — the sweep is many full passes, not one pass over many.
+- **First two verdicts on the record (dogfood):** `RVW-001` — the four-rule "Karpathy" CLAUDE.md →
+  **REJECT** (BOSS already encodes all four as principles + the cohort-aware conscience; a static file
+  would regress toward the frozen-rules brittleness the thread's own top critique names — which is
+  IDEA-014's thesis). `RVW-002` — slaorta's lean/modular CLAUDE.md → **ADAPT** (apply the
+  recency-window to `RESUME.md`'s State section, which duplicates `registry/CHANGELOG.md` and grows
+  unbounded; generalizable shape is a `library/practices/` UP candidate). The REJECT/ADAPT split is
+  the evidence `/vet` routes on merits, not reflex. RVW-001 also surfaced **external confirmation of
+  IDEA-014** (a stranger reasoning to the recalibration thesis) — now cited in that idea.
+
+## 0.40.0 — 2026-06-02
+
+- **`/vet` — the skeptical inbox. The inverse of `/boss-learn`.** From Ajesh's seed: *"if i have new
+  research or best practices, we should have a way where i can just drop it in, and then our mentors
+  and such review and see what we should integrate. reddit is full of best practices, but that doesnt
+  mean all are good ideas."* The last sentence is the whole design.
+  - **Why it's the inverse, not a fork.** `/boss-learn` routes a pattern *you already proved* (built
+    it, it worked, it repeated) UP into `library/` or DOWN into the app — its input has earned trust.
+    `/vet` takes a claim *from a stranger* (a Reddit thread, an HN comment, a blog post, a paper, a
+    "you must do X" tweet) that has earned **nothing**. Its job is the part `/boss-learn` never has to
+    do: decide whether an unproven outside claim deserves to become practice **at all**. ADOPT *hands
+    to* `/boss-learn` (whether → where); it never reimplements it.
+  - **The filter is the product.** A drop folder with no judgment is a bookmark pile. The value is the
+    skeptical read. The skill is **biased toward NO** — most internet best practices don't apply to
+    BOSS, at its stage, for its thesis — and makes a claim *earn* an ADOPT.
+  - **The NO-biased rubric (any one question can sink the claim):** (1) does it contradict a PRINCIPLE?
+    (#6 / `mentor-humane` can veto outright); (2) evidence grade — n=1 vibe vs. pattern-with-data vs.
+    respected practitioner (most claims die here); (3) duplicate or genuinely sharpen?; (4) who does it
+    serve **and harm** (great for `eng-builder`, toxic for `first-product` → ADAPT-with-scoping at
+    best); (5) cost/ceremony (does it make BOSS heavier — R&H #1).
+  - **Four honest verdicts**, mirroring `/extract`'s UP/DOWN/NOT-YET: **ADOPT** (→ `/boss-learn`),
+    **ADAPT** (modified, reasoned), **REJECT — with reason, recorded** (the quietly important one — so
+    the same thread isn't re-litigated next month; the verdict log is BOSS's memory of what it
+    *deliberately didn't* adopt), **NOT-YET** (with a re-open condition). Before vetting, `/vet` reads
+    prior verdicts and won't re-litigate.
+  - **Restraint by design (PRINCIPLE #2):** deliberate-invoke, like `/extract` and `/drift-deep` —
+    **no `vet-loop`, no hook moment, no nudge to "review your inbox."** An automatic research-review
+    obligation would be the ceremony BOSS exists to refuse. It also doesn't *find* research (that's
+    `/deep-research`) — it judges what you bring it.
+  - **Scope: internal-curation first.** `/vet` is a **BOSS-local meta-skill** (lives with `/boss-learn`
+    + `/boss-sync` in `.claude/skills/`, **not** in the founder template) — it vets against
+    `PRINCIPLES.md` + BOSS's own `library/` and routes ADOPT into the BOSS source. The founder-facing
+    version (founder drops a thread → BOSS reads it against *their* canvas/stage/cohort) is the named
+    **UP candidate** (IDEA-016), deferred until the internal version earns it.
+  - Shipped: `.claude/skills/vet/SKILL.md`; drop zone `docs/research/inbox/` + verdict log
+    `docs/research/verdicts/` (each with a README); new **`RVW-NNN`** ID type in `docs/IDS.md`;
+    IDEA-016 captured + the two design forks decided (internal-first; single skeptical pass — the
+    mentor+persona panel is the upgrade if the single pass proves too shallow). Zero-dep held (`npm
+    pack` ships **0** of these — BOSS-local + `docs/`, neither in the `files` allowlist). Gate +
+    judgment suites unchanged (no hook moment, no predicate change).
+
+## 0.39.0 — 2026-06-02
+
+- **`capture` goes judge-backed — the third model-judgment moment, and it ships GRADED from day one.**
+  capture (moment #3, PRINCIPLE #1's own) fired structurally on `≥3 devlog entries + no extraction
+  record` — but a count can't tell a real extraction candidate from three entries of normal in-progress
+  work. That's the exact crude-predicate problem drift (v0.31) and caution (v0.33) already solved with a
+  bounded-read model judgment. v0.39 gives capture the same upgrade.
+  - **The judgment (strictly more restraint — capture can now only fire LESS):** the gate still opens,
+    but before voicing, the model silently reads the ~5 most recent devlog entries and fires ONLY if
+    there's a real candidate — a pattern built **twice** (reusable practice → UP into `library/`), a
+    fix/guard hand-applied in **several places** (hardening → DOWN into core), or a manual **ritual
+    repeated** enough to deserve a skill/loop. If the recent work is one-off distinct features, deep
+    focus on a single still-in-progress thing, or early throwaway spikes — nothing has generalized;
+    **stay silent.** The silent class is trust-critical: nudging `/extract` with nothing to extract
+    earns a NOT-YET every time and trains the founder to tune the conscience out — the premature
+    ceremony PRINCIPLE #2 warns against. Same shape as drift/caution: no model call in the hook, no new
+    state, no predicate change — a bounded-read voicing instruction the model executes in the live turn.
+  - **Shipped:** upgraded `capture` voice frame in the hook lib; `capture` added to `JUDGE_MOMENTS` (so
+    the conscience-frequency ledger logs it as a judge-moment) and to `MOMENT_SIGNALS` (voice-hash
+    source of truth); **`capture.judgment.yml`** (7 labeled cases — 3 should-fire-extractable
+    [practice-twice / guard-in-3-places / repeated-ritual], 3 should-not-fire-nothing-yet
+    [one-off / single-in-progress / spikes], 1 ambiguous [written-twice-maybe]);
+    **`fixtures-devlog-extract.js`** (extractability-focused devlog corpus, distinct from drift's
+    risk-focused one); `replay.js` + `regrade.js` extended with a `capture` row (the MOMENTS registry
+    proves it generalizes — third moment, same engine).
+  - **Graded the free way (per v0.38):** all 7 cases run through isolated reasoning-required Opus 4.8
+    sub-agents; **all 7 agree with the human labels.** `replay.js` reads **GRADED 7/7** for capture
+    (24/24 across drift+caution+capture). Transcripts stamped `generated_via:
+    in-session-subagent-reasoned` + `harness_note`; a real `npm run regrade capture` overwrites them.
+  - Zero-dep held: `npm pack` ships **0** judgment/transcript/extract-fixture files; no `src/` ref.
+    Gate suite **105/0/41** (the predicate is unchanged — `moment-capture.yml` still covers detection).
+    The judgment channel now covers **3 of the conscience's moments**; the remaining structural moments
+    (cost / failure-mode / cost-stale) are binary facts and correctly stay non-judge (a model judge
+    there would be the v0.34 cost trap).
+
+## 0.38.0 — 2026-06-02
+
+- **The conscience's judgment is now MODEL-VERIFIED — `drift` + `caution` read `GRADED 17/17`, not
+  `NEVER_GRADED`. The hole `regrade.js` was built to close (v0.35) is closed — without an API key.**
+  Since v0.32 the judgment surface (`replay.js`) shipped a labeled set + voice-hash tripwire + coverage
+  floors but printed `NEVER_GRADED` loudly: the model had never actually been tested against the labels,
+  so every judge-moment was structurally-checked vibes. Closing it normally needs a paid out-of-band
+  `regrade.js` run (Node `fetch` → Anthropic API). We closed it the free way: `regrade.js` runs two
+  model calls per case *because it executes with no model present* — but a live session **is** Opus 4.8,
+  the same model it would call. Each of the 17 cases (10 drift + 7 caution) was run through an **isolated
+  sub-agent** seeing only the exact voice frame + bounded read the hook injects, and the decisions
+  written as transcripts in `regrade.js`'s own format.
+  - **Result: all 17 decisions agree with the human labels.** The frame and the labels are
+    well-calibrated; the model nails the trust-critical silent class (the on-aim cases where firing
+    would be the false positive that erodes the conscience) — e.g. it reads a missing canvas
+    "Experiment this week" line as a *bookkeeping* gap, not a *validation* gap, when the devlog shows
+    the experiment is already running.
+  - **A real methodology finding, recorded honestly:** a first, terse "output only SILENT or the nudge"
+    harness mislabelled **3 of 17** — one spurious fire (on-aim drift) and two spurious silences
+    (textbook feature-piling / competitor-watching caution cases). Requiring the model to do the
+    "silently read… then judge" reasoning the voice frame *explicitly demands* flipped all three to
+    agree with the label. The lesson: the frame's reasoning instruction is load-bearing, and how you
+    elicit a judgment changes it — exactly the kind of thing the recalibration discipline exists to catch.
+  - **`regrade.js` made importable** — `main()` now runs only on direct invocation; `decisionPrompts`,
+    `MOMENTS`, `loadCases` are exported (so the prompt assembly is reusable/testable and can't drift
+    from the paid path). `--dry-run` still green; importing the module no longer spends.
+  - **Honest provenance, not a masquerade:** every transcript carries `generated_via:
+    in-session-subagent-reasoned` + a `harness_note` stating it was NOT the clean `fetch` harness and
+    that `ANTHROPIC_API_KEY=… npm run regrade` overwrites it as the canonical instrument. The interim
+    grading is real (same model, same frame, same isolation) without overclaiming the provenance.
+  - Zero-dep line held: `npm pack` ships **0** judgment/transcript files; no `src/` reference. Gate
+    suite **105/0/41**; judgment **GRADED 17/17** (was NEVER_GRADED 17). The loud "not yet
+    model-verified" banner is gone.
+
+## 0.37.0 — 2026-06-01
+
+- **`/drift-deep` — the deep, whole-project drift audit. The biggest unused 4.8 lever, now built.**
+  The hook `drift` moment (v0.31) is a cheap always-on tripwire: it reads ~5 recent entries and
+  asks "you named a risk, you're piling work, nothing tests it — is the recent work on-aim?" This
+  is the **deliberate, founder-invoked counterpart** that a bounded read can't do: *read EVERYTHING
+  I've built and tell me, across the whole body of work, whether I'm validating my riskiest bet or
+  building around it.* The 1M-context "am I fooling myself across everything" check — the original
+  finding from the very first 4.8 pass.
+  - **Why a skill, not a hook moment:** a whole-project read can't fire per-prompt — that's the
+    expensive-AI-app trap the v0.34 cost discipline guards against. So the cheap moment stays the
+    everyday tripwire; this is the audit you *invoke* when you want the truth, not a glance. The
+    restraint (no loop, no nudge to "run your audit") is the design — making it a recurring
+    obligation would be the premature ceremony BOSS avoids.
+  - **Broader than the gate** in two ways: it runs even when a validation plan exists (did you
+    *execute* the experiment, or write the plan-line and drift from running it?), and it reads the
+    **actual `src/` code** (what you built is the truest record of what you bet on), not just the
+    devlog tail.
+  - **`/drift-deep` (L1-mvp skill)** — reads the canvas (bet + plan + cells) + ALL devlog + every
+    FEAT spec + `src/` structurally + the ideas; judges each body of work against the bet ("does
+    this *test* the risk or build *around* it?"); reaches a verdict (on-aim / drifting / mixed) with
+    confidence + named gaps + the smallest re-aim; writes `docs/drift-audits/DRIFT-YYYY-MM-DD.md`.
+    Cohort-aware (vibe-virtuoso served most — ships a lot, validates little; domain-expert gets the
+    who-could-be-harmed humane lens on an un-validated risk). Routes back to `/canvas` / `/pretotype`.
+  - **Integration:** the cheap `drift` hook moment now points at `/drift-deep` for the full audit
+    (one terse clause — the cheap nudge stays cheap). Follows the `/extract` precedent — a deliberate
+    skill judgment, tested in use, not by the hook-judgment-eval surface (noted in the skill).
+  - L1-mvp now ships 14 skills. Gate + judgment suites 105/0/41 (the drift voice-hash shifted from
+    the pointer — the tripwire working as designed; no transcripts, so no STALE). The 4.8 leverage
+    arc's last deferred item, landed.
+
+## 0.36.0 — 2026-06-01
+
+- **`boss board` — a live read of what's in flight (IDEA-015, Phase 1).** Occasioned by Ajesh's
+  "internal kanban / fire a html site / Obsidian / almost a Trello board" idea. Convened six advisors
+  (venture, architect, humane, designer + vibe-virtuoso & indie-hacker persona reactions); the result
+  was **unanimous and collapses to one fork: build the *view*, refuse the *app*.** A board BOSS
+  *renders* from state it already holds externalizes the arc for a tired brain; a board BOSS *becomes*
+  (log in, drag cards, keep in sync) is the photo-negative of BOSS and Canvas R&H #1 wearing a UI.
+  **The founder never touches the board — they change the work and it re-renders.**
+  - **`boss board` (new CLI subcommand, [src/board.js](../src/board.js))** — derives four columns
+    (Captured → Taking shape → Building → Shipped) from files that already exist. **Frontmatter is
+    truth — reads each IDEA-*/FEAT-* file's `status`, never `docs/ideas/INDEX.md`** (a hand-maintained
+    table that drifts; a board that trusts a drifting source lies). Pure projection: no
+    `.boss/board.json`, no second source of truth, nothing to sync — so concurrent / out-of-order /
+    agent edits can't corrupt a render (the answer to Ajesh's "picks something out of order" worry is
+    *statelessness*, not merge logic). A promoted idea is represented by its FEAT card (no
+    double-count); blocked FEATs flag `· blocked`.
+  - **Humane constraint honored (mentor-humane override):** the riskiest-assumption status sits
+    *above* the columns — when there's capture but nothing pressure-tested, the evidence line says so
+    plainly and points at `/canvas`. Empty columns are shown, not hidden (the empty cell is the
+    diagnostic). Plain factual copy — no completion-celebration, no gamification, no notifications.
+  - **Deterministic projection, no model in the loop** → it's a CLI verb, not a skill (spending model
+    tokens on `readFile` + string-template would be the anti-pattern the v0.34 frequency-ledger work
+    fought). Ships with the binary → available in every project automatically; **no manifest change**.
+    Lands in IDEA-006's already-portable Layer 1 (zero host contract).
+  - **Earned its keep on first run:** reading BOSS's own repo, it surfaced real INDEX-vs-files drift
+    (IDEA-003 / IDEA-014 are `building` in their frontmatter while INDEX still said `exploring`) — the
+    exact failure mode that justified reading frontmatter over the table.
+  - **Deferred by design (the discipline BOSS preaches, applied to itself):** `--html` (read-only,
+    generate-on-demand, same data model) is gated behind *earn it first* — if `boss board` gets run
+    unprompted each session, build the render; if not, the gate saved the work. Obsidian is mostly
+    documentation (`docs/` is already a vault), not a build. Both captured in IDEA-015 with the
+    write-back caveats. End-to-end tested in `/tmp` (empty / captured-only caution banner / canvas →
+    Taking shape / FEAT supersession / blocked / shipped / placeholder-canvas negative test).
+
+## 0.35.0 — 2026-06-01
+
+- **The recalibration engine — `regrade.js` built + run-ready, and model-recalibration named as a
+  standing discipline (IDEA-014 Phase 1).** Occasioned by Ajesh's direction: adapting to new/
+  different models should be a *standing capability*, not the ad-hoc reaction the v0.31–v0.34 arc
+  was. Picked as the move that *helps BOSS keep improving easily* — because both judge-moments
+  (drift, caution) were `NEVER_GRADED`: the labeled sets existed but the model had never been tested
+  against them. Every future judge-moment would ship as vibes until that closed. `regrade.js` is the
+  keystone that makes all future conscience work *measurable*.
+  - **`regrade.js` (zero-dep, env-gated, out-of-band)** — the calibrator promoted from spec-stub to
+    real harness. Per case: a **decision call** (give the live model the exact voice frame the hook
+    injects + the bounded read the instruction names + a neutral founder turn → fire or stay
+    silent?) + a **judge call** (grade the nudge against the case rubric → structured verdict),
+    writing `transcripts/<moment>/<id>.json` stamped with the current voice-hash. Uses Node built-in
+    `fetch` (no SDK). **`--dry-run`** verifies the whole pipeline (prompt assembly across all 17
+    cases + the judge parser) with no API and no spend — verified green; the live `fetch` is the
+    only unexercised line (a standard POST). `npm run regrade`; `BOSS_REGRADE_MODEL` to point at a
+    different model.
+  - **`moments.js`** — shared voice-hash source of truth so `replay.js` and `regrade.js` *cannot*
+    disagree on the fingerprint (a mismatch would mark every transcript STALE forever). `replay.js`
+    refactored onto it (same hashes; suites regression-clean).
+  - **`docs/architecture/MODEL-RECALIBRATION.md`** — the standing checklist (IDEA-014's earned
+    slice). Two triggers: a new same-vendor model (*leverage more* — re-grade, revisit each
+    boundary for "can this move UP now?", read the frequency ledger, check context headroom) and a
+    new host running a different model (*degrade gracefully* — judge-moments fall back to predicate-
+    only). `regrade.js` is its engine; re-running it on a new model **is** recalibration. Named as a
+    PRINCIPLE-#1 **UP** candidate (founders face the same model-migration problem — `/claude-api`
+    already migrates their apps).
+  - **Deferred, loudly:** the per-host model-capability profile + a `/recalibrate` skill — until a
+    second model/host exists (a `/recalibrate` with nothing to recalibrate *to* is the premature
+    ceremony v0.34 dodged). Token accounting stays host-contract territory (IDEA-006).
+  - Gate + judgment suites 105/0/41; `npm pack` ships 0 tooling files. The judgment is now
+    *run-ready* to become model-verified — one `ANTHROPIC_API_KEY=… npm run regrade` away (no key in
+    the build env, so the live grade is the founder's to trigger).
+
+## 0.34.0 — 2026-06-01
+
+- **Conscience frequency ledger — BOSS eats its own `/ai-cost` dogfood, honestly. The last build
+  of the 4.8 arc.** As judge-moments multiplied (v0.31 drift, v0.33 caution now do model judgment
+  in the live turn), the conscience began costing real tokens per prompt — while BOSS preaches
+  cost discipline and never measured its own. The first 4.8 pass flagged the trap: *BOSS becomes
+  the expensive-AI app it warns against.* (IDEA-013.)
+  - **The reframe is the decision (mentor-architect):** the build started as "conscience *cost*
+    instrumentation" and was reframed to **frequency — facts, not estimates.** The hook never
+    calls a model, so a char→token estimate would manufacture a billable-looking number while
+    blind to its dominant term (the induced bounded reads judge-moments trigger in the main turn).
+    That's *lying with numbers* — the exact cost-theater BOSS warns against; PRINCIPLE #2 vetoes
+    it as premature ceremony. The *real* problem under the cost framing is **over-firing** — the
+    actual way a conscience becomes costly/annoying, and a number you'd act on. So: measure
+    frequency honestly; defer the token question to where honest token data lives (host-side,
+    IDEA-006).
+  - **`.boss/conscience-log.jsonl`** (gitignored) — the hook appends one line per fire:
+    `{ts, moments[{moment,confidence}], judge(bool), injected_chars, cohort}`. Facts only — **no
+    token/dollar estimate.** A **separate BOSS-meta ledger**, never the founder's
+    `.boss/cost-log.jsonl` (BOSS's overhead ≠ the founder's app cost).
+  - **First correctness-invisible fire-path side effect.** The hook goes from pure-emit to
+    has-a-side-effect; `logActivity` runs only past the silent early-exit, append-only, single
+    write, in its own swallowing try/catch. **Verified byte-identical** hook output with and
+    without the ledger writable — delete it and the conscience behaves the same.
+  - **`boss conscience activity`** (alias `cost`, which prints the honest frequency-not-tokens
+    reframe) — fires, judge-moment share, median injected chars, per-moment counts, and the
+    **over-fire smell** (clustering: a moment firing ≥4×/hour or ≥8×/24h — flagged because no
+    per-prompt denominator exists without logging non-fires, which would break the instant
+    property). Plus a one-line activity summary in `boss status --conscience`.
+  - **Measure-only; self-throttle deferred indefinitely.** A throttle would gag the conscience
+    exactly when a drifting founder generates more prompts and needs it most — **humane before
+    viable**, and a one-way door (every "why didn't it speak?" becomes unfalsifiable). This ledger
+    is the *evidence* that would earn that conversation, not a step toward it. Token/dollar
+    estimation also deferred — host-contract territory (the only honest token count comes from the
+    host, not the hook). `JUDGE_MOMENTS` set added to the hook lib. Gate + judgment suites
+    regression-clean (105/0/41; drift + caution covered).
+
+## 0.33.0 — 2026-06-01
+
+- **`caution` goes judge-backed — depth vs. avoidance, and the first reuse of the v0.32 judgment
+  machinery on an *existing* moment.** Moment #1 (the conscience's flagship "what does this
+  prove?") fires when ≥3 captures exist with no filled riskiest assumption. But the predicate
+  counts *total* captures — it can't tell **depth** (one idea getting sharper, converging toward a
+  canvas) from **avoidance** (capturing-lots / validating-nothing). That ambiguity is `m1-snf-021`,
+  which the gate runner has *skipped since v0.16* because no predicate can make the call. v0.33
+  makes caution judge-backed and resolves it.
+  - **Voice-frame upgrade** (`signalAsContext`, `caution` branch): the gate still opens, but the
+    model now silently reads the active idea's capture log and judges before voicing — *one idea
+    sharpening (narrowing the user, finding the real pain, wrestling the same question) = stay
+    silent; idea-hopping / feature-piling / market-notes-without-a-bet = fire and name the specific
+    pattern.* This is **strictly more restraint** — caution can now only fire *less*, never more.
+  - **First reuse of the judgment surface (v0.32) on an existing moment** — proves the machinery
+    isn't drift-specific. `replay.js` is now **multi-moment**: a `MOMENTS` registry (drift +
+    caution) drives one shared engine (per-moment voice-hash tripwire + well-formedness + coverage
+    + grading status). Adding a judgment moment = one registry row + a `<moment>.judgment.yml`.
+  - **`caution.judgment.yml`** — 7 labeled cases: 3 should-fire-avoidance (idea-hopping,
+    feature-piling, market-notes), 3 should-not-fire-depth (the m1-snf-021 case made concrete —
+    narrowing, wrestling, converging), 1 ambiguous (depth-with-a-detour, `acceptable:
+    [fires, silent]`). Content-rich capture-log fixtures (`fixtures-capturelog.js`) — the prose
+    matters because the judgment is semantic.
+  - **m1-snf-021 closed at the right layer:** the gate runner keeps skipping it (correctly — no
+    predicate can make the call), now annotated "RESOLVED in v0.33; covered by the judgment
+    surface," not "unimplemented." The case the gate *couldn't* test is exactly what the judgment
+    channel was built for — the clearest proof v0.32 earned its keep.
+  - Gate suite regression-clean **105/0/41**; both judgment moments well-formed + covered; caution
+    voice-hash updated (tripwire reads the live frame). End-to-end tested via `boss new`: caution
+    fires on scattered captures and ships the new depth-vs-avoidance instruction. The judgment
+    remains **not yet model-verified** (replay prints NEVER_GRADED loudly) — first STALE builds
+    `regrade.js`.
+
+## 0.32.0 — 2026-06-01
+
+- **Judgment-quality eval channel — closing the hole `drift-loop` (v0.31) opened.** drift was the
+  first moment whose *detection is a model judgment*, not a predicate. The existing eval runner
+  tests only the **gate** (does the hook fire on the right structural state); it stops at the
+  door. Whether the model correctly calls drift-vs-on-aim *past* the open gate was unevaluated —
+  a named crack in the "no moment ships without evals" floor, in exactly the spot 4.8 made
+  load-bearing (`eval.md`: shipping detection logic with no way to know if it's right = "vibes-
+  based AI in BOSS"). v0.32 builds the complement.
+  - **The method (mentor-architect pass):** two surfaces, two cadences, on purpose. The gate-eval
+    (`runner.js`, every commit, $0) stays. The **judgment surface** (`conscience-evals/judgment/`)
+    is **golden transcripts gated by a voice-hash tripwire** — *not* pure LLM-as-judge (breaks the
+    free/deterministic/CI property) and *not* pure golden transcripts (rot silently when the voice
+    frame changes). Golden transcripts are the committed dataset; the tripwire makes their
+    staleness *loud*; LLM-as-judge regenerates them out-of-band.
+  - **`judgment/replay.js` — zero-dep, every commit.** (1) well-formedness — every case is a
+    genuine open-gate state (filled risk, devlog ≥3 entries, no experiment line) with a coherent
+    label; (2) **voice-hash tripwire** — fingerprints the exact `drift` instruction the model runs
+    (`composeContext` for a drift signal); a transcript recorded against a different hash is
+    `STALE` and replay says so loudly (golden transcripts that can't detect their own staleness
+    are an eval that lies); (3) coverage floors (no silent caps; the on-aim/should-not-fire class
+    is trust-critical and meant to *grow* toward ≥10); (4) grading status `GRADED/STALE/
+    NEVER_GRADED/REGRESSION`. Exit 1 on malformed/coverage/regression; exit 0 + loud warnings on
+    never-graded/stale. **All four states proven** (GRADED/STALE/REGRESSION/NEVER_GRADED tested).
+  - **`judgment/drift.judgment.yml` — the labeled ground truth (Husain: build the set first).**
+    10 cases: 4 should-fire-and-name-gap (incl. the sharpest — building the integration the risk
+    explicitly *defers*), 5 should-not-fire-on-aim (the trust-critical class, incl. the hard
+    "on-aim but informal — no experiment line, but the work IS the test" case), 1 ambiguous
+    (`acceptable: [fires, silent]` so the grader doesn't punish a defensible call). Content-rich
+    paired devlog fixtures (`fixtures-devlog.js`) — the *prose* matters because the judgment is
+    semantic. `must_reference`/`must_not` rubric per case for the future LLM-judge.
+  - **`judgment/regrade.js` — paid, out-of-band, DEFERRED by design.** The calibrator (decision
+    call → judge call → write transcript stamped with the current voice-hash). Per the architect's
+    staged cut: ship the labeled set + replay + tripwire first; build the API half *the week the
+    first STALE tripwire fires*. Runnable spec-stub documents the full algorithm + env-checks. When
+    built: zero-dep (Node `fetch`, no SDK), env-gated on `ANTHROPIC_API_KEY`, never on the commit
+    path, never imported by `src/`, never in the `files` allowlist (lives under `docs/`).
+  - **The zero-dep line, pinned:** the rule is *shipped surface (`src/`, `files`) stays
+    dependency-free* — not "dev tooling can't call a model." Confirmed: `npm pack --dry-run` ships
+    **0** judgment files; no `src/`/`bin/` reference to the eval dirs.
+  - **Dogfooded in `eval.md`:** a model-judgment moment cannot ship its detection with only a
+    gate-eval — the drift signal + exit predicates now require a judgment set + replay coverage for
+    `drift` and every successor. Shared YAML parser extracted to `conscience-evals/lib/yaml-eval.js`
+    (gate runner + replay use one copy; gate suite regression-clean at 105/0/41). New npm scripts:
+    `eval:gate`, `eval:judgment`, `eval`.
+  - **Honest scope:** v0.32 ships the labeled set + the staleness machinery + coverage discipline.
+    The judgment is **not yet model-verified** — replay prints `NEVER_GRADED` loudly so a green run
+    is never mistaken for a graded judgment. The first `STALE` is the trigger to build `regrade.js`.
+
+## 0.31.0 — 2026-06-01
+
+- **`drift-loop` — the closest loop to why BOSS exists, and the first moment that fronts a
+  *model judgment* the predicate can't make.** PRINCIPLES.md opens with *"build faster without
+  fooling themselves… continuously checking the work against real pain, real buyers."* For 30
+  releases the conscience caught *structural* gaps (no canvas, no budget, no failure-states)
+  but never the gap it names first: **you named the bet that could sink this, then spent your
+  sessions building something else.** v0.31 closes it — occasioned by a "what does Opus 4.8
+  change about BOSS" pass whose load-bearing finding was *the hook=detection / model=voice
+  boundary has moved*: a stronger model can do the semantic judgment regex can't, so the
+  conscience's first real "are you fooling yourself" check becomes buildable.
+  - **`drift-loop` (L1-mvp, hook-runner)** — fires the new `drift` moment when (a) a canvas has
+    a real **Riskiest assumption** (same predicate `canvas-loop` / `spec-loop` use) AND (b)
+    `docs/devlog.md` has ≥3 dated entries (work accumulating, same threshold as
+    `extraction-loop`) AND (c) no real **Experiment this week** validation plan exists yet.
+    Sits in the gap *between* `caution` (no risk named) and the `done` graduation (risk has a
+    plan). Confidence scales on devlog overshoot: 3 → low, 4–5 → medium, 6+ → high.
+  - **The architecture, pinned (mentor-architect pass):** this is **not a new detector** — it's
+    a *predicate-gated, bounded-read voicing instruction on the existing `UserPromptSubmit` →
+    `additionalContext` channel*. The cheap Node predicate is the gate (and the cost control);
+    the model does the stated-vs-actual comparison *in the live turn*, reading a **bounded** set
+    (riskiest-assumption line + ~5 recent devlog entries + the open FEAT — never the whole
+    project). No model call in the hook, **no new host primitive** (IDEA-006 contract untouched),
+    **no new state**, no new predicate vocabulary. Same shape as `extraction-loop`/`/extract`,
+    but **hook-fired not skill-invoked** — because a founder who has drifted from their own
+    stated bet is exactly the founder who won't think to *ask* whether they've drifted.
+  - **New `drift` voice frame** in `signalAsContext` (loop-runtime.js): instructs silent bounded
+    read → judge → name the *specific* gap ("you said X is the bet; the last sessions built Y, Z;
+    neither tests X") → point at `/canvas` or `/pretotype`. **Stay silent when on-aim** (silence
+    is the correct output). Not a "you've been productive!" reward, not a generic "you should
+    validate" line — the value is the specific comparison. Cohort-aware (returning-founder gets
+    the harder conviction cut; first-product gets "test your riskiest bet" taught plainly;
+    domain-expert gets the who-could-be-harmed lens on the named risk).
+  - **Eval coverage from the start** — `moment-drift.yml` (14 cases: 4 should-fire incl. the
+    placeholder-experiment edge + the drift/capture co-fire; 10 should-not incl. plan-recorded,
+    under-threshold, unnamed-risk, dropped-idea, pre-MVP, and the documented `any_file_matches`
+    masking limitation). Runner extended: `buildCanvasFile` now writes the experiment line.
+    Suite **105/0/41** (up from 91). *No moment ships without evals* — held. End-to-end tested
+    in `/tmp`: fires low→high as devlog grows, goes silent when the validation plan is recorded.
+  - L1-mvp now ships 12 skills + **8 loops**. Honest scope note: hook evals test the *gate*;
+    judgment-quality (does the model correctly call drift vs. on-aim) is the model's layer,
+    tested the way `/extract`'s judgment is — not this runner.
+
+## 0.30.0 — 2026-05-27
+
+- **Closing the two remaining audit gaps — cost-log read cadence + failure-state stub
+  loophole.** Same shape as v0.27: no new product axis; the discipline tightens around
+  surface already shipped. The two gaps the audit named after v0.26 — *"the weekly review
+  cadence is unenforced"* and *"failure-state handlers can be stubs forever"* — were real
+  holes. v0.30 closes both.
+  - **`/cost-review` skill (L1-mvp)** — reads `.boss/cost-log.jsonl`, aggregates by FEAT +
+    user + cohort, compares against `docs/ai-cost-budget.md`, flags surprises (cost
+    outliers > 10× median; user outliers; FEAT skew; quiet upward drift), writes
+    `docs/cost-reviews/REVIEW-YYYY-MM-DD.md` with headline / numbers / variance / surprises /
+    actions / next-review fields. Cohort-aware framing (indie-hacker gets %-of-revenue;
+    returning-founder gets unit economics; domain-expert gets privacy-first confirmation
+    before showing any review data). Surfaces mentor handoffs (architect for cost-shape;
+    business for unit-economics) when overages exceed 10% of monthly cap.
+  - **`cost-review-loop` (L1-mvp, hook-runner)** — **second time-of-work entry pattern**
+    (after extraction-loop). Entry: `docs/ai-cost-budget.md` exists (the deontic moment —
+    declaring the budget commits you to reading it). Exit: ≥1 file in `docs/cost-reviews/
+    REVIEW-*.md` with a `^- \*\*Total spend:\*\*` line. The first review closes the loop;
+    recurring re-opening waits on time-aware predicate vocabulary (same dependency as
+    extraction-loop). New **`cost-stale` moment** added to `signalAsContext` voice frame.
+  - **`/ai-failure-states` template — Eval-tested column added.** Each failure-state row
+    (garbage / refusal / hallucination / timeout / cost-spike) now carries an **Eval-tested**
+    field naming which eval case actually exercises the handler, OR marked **STUB** with an
+    override required per IDEA-008. *Closes the stub-forever loophole at the declaration
+    layer.* Voice note: stubs are still legitimate; *stubs without an override-with-re-open-
+    condition* are not.
+  - **`/evals` skill — failure-mode coverage requirement.** For AI-mediated FEATs, the eval
+    set must include at least one `should-fail` case per declared failure state, categorized
+    by `failure_mode` matching the canonical names (`garbage`/`refusal`/`hallucination`/
+    `timeout`/`cost-spike`/`other`). *Closes the stub-forever loophole at the test layer.*
+    The two layers (declaration + test) compose: handler stubs in code → declaration in
+    docs → eval cases that exercise the declarations → real coverage.
+- **Eval coverage from the start (v0.27 rule held).** `moment-cost-stale.yml` ships with 9
+  cases: should-fire (no review; empty directory; stub file without canonical line); should-
+  fire-multi (cost + cost-stale together when logger isn't wired but budget is); should-not-
+  fire (full closure with review on record; pre-budget projects; LLM calls without budget
+  declared — confirms entry is gated on budget doc, not LLM calls; Quickstart-shape projects).
+  3 existing "full discipline declared" cases (m-cost-101, m-fm-101, m-cap-202) updated to
+  also satisfy cost-review-loop's exit predicate, since they assert all-loops-closed.
+- **Suite count: 91 passed / 0 failed / 41 skipped (133 loaded)** — up from 83/0/41. Every
+  hook-emitted moment now has eval coverage. The "no moment ships without evals" discipline
+  is the new floor.
+- **What v0.30 does NOT do:** auto-enforce the failure-mode eval coverage (the `/evals`
+  skill names the requirement; nothing fails the build if the eval set lacks coverage; that
+  would be a v0.31+ tightening if founders ship handlers-without-evals systematically);
+  recurring cost-review-loop (needs time-aware predicate vocabulary); pull `.boss/cost-
+  log.jsonl` into the skill review output automatically (the parsing logic lives in the
+  skill's instructions; Claude does the reading per-call to keep zero-dep CLI).
+- **The audit is closed.** The three load-bearing gaps from the post-v0.26 audit (eval
+  coverage, /welcome onboarding, moment #3 capture) shipped in v0.27 / v0.28 / v0.29; the
+  two discipline-hole gaps in already-shipped surface (cost-review cadence, failure-state
+  stub loophole) shipped in v0.30. The right-to-defer items (brownfield, /consult, AI-first
+  archetype template, IDEA-003 finish, etc.) remain on the v0.31+ backlog. Audit-driven
+  hardening complete; the next move is feature work.
+
+## 0.29.0 — 2026-05-27
+
+- **Moment #3 lands — PRINCIPLE #1's own discipline, encoded (finally).** For 28 releases the
+  conscience surfaced 6 moments (caution / done / restraint / coherence / cost / failure-mode)
+  but never one for PRINCIPLE #1 itself — the *pause to extract patterns UP or DOWN* rule that
+  defines what BOSS is. *"Capture"* was deferred 5 consecutive releases as "needs LLM-as-judge
+  or heuristic design." v0.29 ships the heuristic-plus-judgment version: a hook-runner loop
+  on a simple devlog heuristic + a skill that does the real judging.
+  - **`extraction-loop` (L1-mvp, hook-runner)** — entry: `docs/devlog.md` has ≥3 dated
+    entries (regex `^## \d{4}-\d{2}-\d{2}`); exit: ≥1 `docs/extractions/EXTR-*.md` file has
+    a `- **Route:**` line. **First hook-runner loop whose entry is *time-of-work* (devlog
+    count) rather than *file-state predicate*.** Sets the precedent for future judgment-
+    required moments that can't be detected by file regex alone. Threshold of 3 mirrors
+    PRINCIPLE #1's *"the third time the same work repeats"* signal.
+  - **`/extract` skill (L1-mvp)** — the LLM-as-judge counterpart. Reads recent commits, the
+    last 3-5 devlog entries, the current FEAT, `library/`, and `src/`. Identifies 1-3
+    candidate patterns by three signals (Signal A: same work repeated; Signal B: named-and-
+    stable shape; Signal C: load-bearing decision). Routes each candidate **UP** (into BOSS's
+    `library/<cat>/` via `boss learn`) or **DOWN** (into the app's `src/` as a named refactor
+    target) or honest **NOT-YET** (the third legitimate answer; recording the *not yet* IS
+    the discipline). Writes `docs/extractions/EXTR-NNN-<slug>.md` with full routing record +
+    "what didn't make the cut" section. Cohort-aware (returning-founder gets the seasoned
+    `"what did you do twice?"` prompt; first-product gets gentler framing; domain-expert
+    leans NOT-YET-with-caveats for regulated logic).
+  - **New `capture` moment in the conscience** — added to `signalAsContext` in
+    `lib/loop-runtime.js`. Voice frame names the inflection in plain language; explicitly
+    rejects *"you've been productive!"* reward framing (PRINCIPLE #1 is the discipline, not
+    the dopamine). Cohort decides the specific framing.
+  - **`EXTR-NNN` ID type added** — `docs/IDS.md` template updates to name it under MVP-mode
+    unlocks alongside `FEAT-NNN` / `FIX-NNN` / `BUG-NNN`.
+  - **L1-mvp manifest** — now ships 11 skills + 6 loops. `claude-append.md` names `/extract`
+    and `extraction-loop` alongside the others; calls out the *"two destinations, not one"*
+    framing per PRINCIPLE #1.
+- **Eval coverage from the start (closes the v0.27 hole going forward).** `moment-capture.yml`
+  ships with 11 cases: should-fire (3, 5 entries; empty extractions/ dir doesn't close); should-
+  not-fire (≤2 entries; closed by UP route; closed by NOT-YET route; empty project; empty
+  devlog file); Quickstart-shape sanity (no devlog → no fire); multi-loop isolation (all four
+  AI-first/extraction loops closed independently). The discipline applied: **no moment ships
+  without its evals.**
+- **Runner upgrades** — new fixtures: `devlog_3_entries`, `devlog_2_entries`, `devlog_5_entries`,
+  `extraction_record_up`, `extraction_record_not_yet`. Loads `moment-capture.yml` in main().
+  **Suite count: 83 passed / 0 failed / 41 skipped (124 loaded)** — up from 73/0/41.
+- **What v0.29 does NOT do:** auto-recurring extraction-loop (closes after one extraction;
+  re-opening requires founder action or future predicate vocabulary with time-aware checks
+  like *"N devlog entries since last extraction"*); execute the actual extraction (the skill
+  proposes UP/DOWN routes; `boss learn` handles UP execution; refactors are the founder's
+  call); LLM-call-out detector (some "is this reusable?" judgments would benefit from an
+  external model call, but Claude-running-the-skill IS the LLM-as-judge — keeping it in-
+  context is simpler and cheaper).
+- **The architectural precedent.** extraction-loop is the **first time-of-work entry**
+  pattern. Future loops that need heuristics like "X has accumulated since Y" can reuse this
+  shape (count files of one type; gate by absence of files of another type). Documented in
+  the loop spec for future authors. Tightens the v0.18 runtime's vocabulary without changing
+  the runner.
+
+## 0.28.0 — 2026-05-27
+
+- **`/welcome` — the gentle first-run orientation (closes the v0.19 cohort gap).** The v0.19
+  personas pass surfaced that `first-product`, `vibe-coder-newbie`, and `non-tech-founder`
+  cohorts bounce off BOSS because the first thing they see is *"drop your PRD."* Power-user
+  framing, beginner cohort. v0.28 adds the gentler door: a Quickstart skill that orients in
+  ~1 minute, defines terms inline, names the conscience as a *nudge never a block*, surfaces
+  override + pause as discoverable affordances, and routes to `/boss` or `/triage` based on
+  what the founder is ready for.
+- **Cohort-aware depth** — the skill branches on `.boss/config.json` cohort:
+  - `first-product` / `vibe-coder-newbie` / `non-tech-founder` / null → **full tour**: what
+    BOSS is, what's in the folder, what to do next, how the conscience works, how modes level
+    up, where to find help. Each section 2-3 sentences; terms defined inline; plain language.
+  - `eng-builder` / `vibe-virtuoso` / `indie-hacker` / `returning-founder` → **30-second
+    version**: one paragraph, pointer at `/boss`. *"You probably don't need the tour."* Then
+    stops; doesn't elaborate.
+  - `domain-expert` → **middle path**: full tour with high-stakes framing inline (privacy-
+    first defaults, human-in-the-loop on hallucination, conscience errs toward speaking).
+- **Discoverability fix for IDEA-011** — the skill explicitly surfaces `boss conscience pause`
+  + the override grammar (*"deviation conscious, recorded, never blocked, never forgotten"*).
+  Previously these were documented in CHANGELOG / docs/ideas; now they're surfaced in the
+  cohort's first 5 minutes with BOSS. Partial close on IDEA-011 Phase 2's override-
+  discoverability sub-gap.
+- **Cohort question moved upstream** — `/welcome` asks the cohort question first (when unset),
+  before any orientation; `/boss` step 6 stays as the backup path for founders who go
+  directly to `/boss`. The cohort persists in `.boss/config.json` for the conscience hook.
+- **`boss new` output updated** — the post-scaffold "Next:" block now names BOTH paths:
+  `/welcome` for the first-time door, `/boss <idea|PRD>` for the power-user door. Same
+  language as CLAUDE.md (template) top.
+- **L0-quickstart manifest** — adds `welcome` to the skills list (now 6 skills). No new
+  agents, no new loops, no new hooks. Quickstart still ships a tiny agentic footprint.
+- **L0 CLAUDE.md (template)** — adds a top-line nudge: *"First time? Run `/welcome` — gentle
+  orientation, takes a minute, defines terms inline. Already familiar with BOSS? Skip to
+  `/boss <your idea or PRD path>` to spin up."* The first thing a fresh founder reads.
+- **End-to-end tested in /tmp** — `boss new welcome-test` → 6 skills land including welcome;
+  CLAUDE.md surfaces /welcome at top; 73/73 conscience evals still pass.
+- **What v0.28 does NOT do:** auto-trigger `/welcome` on first Claude prompt (the CLAUDE.md
+  surfacing is enough; auto-trigger via hook would conflict with cohort `eng-builder` /
+  `returning-founder` who explicitly *don't* want it); ship an end-user onboarding flow (that
+  was named in IDEA-012 as a separate v0.27+ candidate; this skill is founder-onboarding,
+  not end-user-onboarding); add a `/welcome` for V1 / Scale modes (the lift is once, not per-
+  mode — re-running in MVP+ is supported but the orientation doesn't change much).
+
+## 0.27.0 — 2026-05-24
+
+- **Conscience evals — closing the discipline-on-the-discipline-tool hole.** Four moments
+  (restraint / coherence / cost / failure-mode) had shipped without eval coverage; the brake
+  the discipline named (*"43/43 pass"*) was eroding silently. v0.27 closes the three
+  hook-emitted ones; restraint is skill-side and stays out of the hook eval suite by design.
+  - **`moment-cost.yml`** — 12 cases covering: Anthropic / OpenAI / Vercel AI SDK
+    (generateText, streamText) entry detection; partial closure (budget doc alone, logger ref
+    alone); full closure; multi-moment co-firing with failure-mode at the first-LLM-call
+    inflection; should-not-fire on no-LLM and empty projects.
+  - **`moment-failure-mode.yml`** — 8 cases isolating the failure-mode signal (close the cost
+    loop to single out failure-mode); covers partial closure (doc alone, handlers alone),
+    full closure, and the no-entry case.
+  - **`moment-coherence.yml`** — 10 cases covering: className= / styled-components / inline
+    style={} entry detection; confidence scaling (3 declarations = low; 12+ = high); partial
+    closure (tokens doc alone, refs alone); full closure; under-threshold (2 declarations)
+    and no-UI cases.
+  - **Runner upgrades:**
+    - Loads both L0-quickstart AND L1-mvp loops (L1 loops live in `stages/L1-mvp/template/
+      docs/loops/`; previous runner only loaded L0, which is why three moments could ship
+      without coverage).
+    - **`src_files` / `docs_files` / `other_files` in `project_state`** — materializes
+      arbitrary file paths via the new `FIXTURES` registry (`runner.js`). The minimal YAML
+      parser doesn't support multi-line scalars, so file content lives in the runner as
+      named fixtures: `anthropic_call`, `openai_call`, `vercel_ai_call`, `vercel_ai_stream`,
+      `cost_budget_doc`, `cost_logger_ref`, `failure_states_doc`, `failure_handlers_ref`,
+      `style_decls_low/high/two`, `style_styled_components`, `style_inline_objects`,
+      `design_tokens_doc`, `token_refs`, `no_llm_code`, `empty`.
+    - **Multi-moment assertions** — `expected_detection.moments: [cost, failure-mode]`
+      asserts set inclusion across the actual signals list (not just the first signal).
+      Useful for cases where multiple loops fire simultaneously by design.
+    - **Single-moment fallback** — when `expected_detection.moment` is specified, the
+      assertion also checks the moments list (not just the first signal), since signal
+      ordering is filesystem-dependent.
+  - **README updated** — current-cut section names all five tested moments; documents the
+    multi-moment format and the FIXTURES registry pattern.
+- **Suite count: 73 passed / 0 failed / 41 skipped (114 loaded)** — up from 43/43. The 41
+  skipped are unchanged: 20 moment-2 cases that live in `/canvas` skill prompt (no hook
+  detector), 6 signal-text evals (separate runner), 15 cases gated on suppress_if / devlog
+  awareness / session-state tracking that aren't yet implemented.
+- **What v0.27 does NOT do:** sharpen the first-cut wording in any YAML (Ajesh's sharpening
+  pass remains pending — flagged in each file's frontmatter); ship a skill-eval runner for
+  `spec-loop` / `pretotype-loop` (those are skill-side; need a different runner pattern);
+  add a sixth moment for "capture" (PRINCIPLE #1 — gap #1 still on the table; needs an
+  LLM-as-judge or heuristic detector design).
+- **Why this isn't a feature release:** no skills, agents, loops, or hooks shipped. The
+  discipline tightened around what's already shipped. Same shape as v0.24 (positioning
+  pass): not the most exciting release, plausibly the highest-leverage one this stretch.
+
+## 0.26.0 — 2026-05-24
+
+- **AI-first product template — BOSS's home turf, made first-class.** The v0.24 positioning
+  named BOSS as *"the thinking layer for AI-native founders."* v0.26 ships the concrete
+  artifact that earns the name: a conductor skill that walks the founder through the AI-first
+  discipline **from day one**, plus the missing piece (failure-states design) that completes
+  the spine. Cursor + a folder is "AI-native." BOSS + `/ai-first-init` is *"AI-first with
+  discipline from day one."*
+  - **`/ai-first-init` skill (L1-mvp)** — the **conductor**. Walks the founder through five
+    steps: (1) declare what's AI-mediated → `docs/ai-first.md`; (2) seed structured outputs
+    (Liu) → `docs/schemas/`; (3) seed eval set early (Husain) → `/evals --new`; (4) declare
+    cost budget upfront → `/ai-cost`; (5) design failure states → `/ai-failure-states`. The
+    "from day one" framing: declare the discipline *before* the first AI-mediated FEAT ships,
+    not after the first bill / hallucination / refusal.
+  - **`/ai-failure-states` skill (L1-mvp)** — the **missing piece**. Walks the founder through
+    five guaranteed failure modes (garbage / refusal / hallucination / timeout / cost-spike),
+    each with a project-specific declared response + stub fallback handler in code (so the
+    discipline is wired before the founder forgets). Writes `docs/ai-failure-states.md`.
+    Cohort-aware: `first-product` gets named patterns; `eng-builder` gets the unhandled-path
+    angle; **`domain-expert` defaults to human-in-the-loop on hallucination** (no retry-loop
+    on medical/legal/financial output).
+  - **`ai-failure-state-loop` (L1-mvp, hook-runner)** — entry: ≥1 LLM SDK call site
+    (parity with `cost-budget-loop` — the two failure modes always coexist at the
+    AI-mediated boundary). Exit: `docs/ai-failure-states.md` exists AND code references at
+    least one fallback handler (`handleGarbageResponse` / `handleRefusal` /
+    `handleHallucination` / `handleTimeout` / `handleCostSpike` or snake_case Python
+    equivalents). New `failure-mode` moment added to `signalAsContext` voice frame.
+  - **`/spec` upgrade** — for AI-mediated FEATs, the spec template now includes a **Failure
+    states** section alongside the existing evals + validated-learning fields. Names which
+    of the five failure states the FEAT must handle + which fallback handler it routes to.
+    Acceptance criteria are asked to reference at least one failure-state path.
+  - **`/boss` AI-native nudge (L0-quickstart)** — when the founder names the model as
+    load-bearing during spin-up (the product doesn't work without it), `/boss` names it
+    explicitly back and recommends *"after `boss unlock mvp`, run `/ai-first-init`."* The
+    recommendation is the artifact; `/boss` never runs it for the founder.
+- **`docs/ai-first.md` as the declaration contract.** Future FEATs read this doc; future
+  `/spec` runs reference its cross-reference fields. If the doc says "deterministic" for a
+  feature and a PR puts an LLM call in there, that's a real change worth a re-spec —
+  promotes design-by-archeology into design-by-declaration.
+- **L1-mvp manifest now ships 10 skills + 5 loops + 4 agents.** `claude-append.md` names
+  `/ai-first-init`, `/ai-failure-states`, and `ai-failure-state-loop` alongside the others.
+- **The lineage cited.** Husain (failure-mode categorization extended from evals → UX), Liu
+  (structured outputs as the contract that makes failure-detection mechanical), Karpathy
+  (the failure surface IS the design surface — designing the happy path is the easy 20%),
+  Mollick (cost-as-design-input continuing from v0.25). No new mentors added; the discipline
+  is the existing roster applied to the AI-mediated boundary.
+- **Conscience regression-clean.** The new loop lives in L1-mvp; the eval runner only loads
+  L0-quickstart loops; no eval fixtures fire `failure-mode`. End-to-end tested in `/tmp`:
+  fresh project → `boss unlock mvp` → drop LLM SDK call in `src/` → hook fires BOTH `cost`
+  AND `failure-mode` simultaneously (the two loops share entry but track different exits).
+
+## 0.25.0 — 2026-05-24
+
+- **AI cost discipline — the universal-cohort feature lands.** Per IDEA-012's persona overlay,
+  AI cost was the only candidate every cohort cared about. Now it's first-class in MVP mode:
+  the founder gets nudged to declare the bill at the first LLM SDK call, not after the first
+  surprise invoice.
+  - **`/ai-cost` skill (L1-mvp)** — walks the founder through declaring `docs/ai-cost-budget.md`
+    (per-user/day + monthly cap + model rationale + review cadence + breach grammar), suggests
+    a ~30-line cost-logger wrapper (TypeScript + Python examples) writing to
+    `.boss/cost-log.jsonl`, and surfaces mentor handoffs (`mentor-architect` for cost-shape →
+    architecture; `mentor-business` for cost-per-user → pricing). **Cohort-aware defaults:**
+    `first-product` $5/user/day strict; `vibe-virtuoso` inspect-only; `eng-builder` BYO;
+    `indie-hacker` cost-as-%-of-revenue framing; `domain-expert` $20/user/day + **privacy-
+    first logging (no PII, no prompt body)** + regulatory caveats; etc. The founder edits to
+    fit the bet; the cohort sets the starting frame.
+  - **`cost-budget-loop` (L1-mvp, hook-runner)** — entry: `src/**` contains ≥1 LLM SDK call
+    site (regex covers `anthropic` / `@anthropic-ai/sdk` / `openai` / `OpenAI(` / `Anthropic(` /
+    `messages.create` / `chat.completions.create` / Vercel AI SDK `generateText` /
+    `streamText`); exit: `docs/ai-cost-budget.md` exists AND code references the logger
+    wrapper (≥1 occurrence). Threshold of 1 (not 3 like design-tokens-loop) because cost
+    discipline is **deontic at the first call** — there's no "exploratory" version of token
+    spend. The first call hits a real billing meter.
+  - **New `cost` moment in the conscience** — added to `signalAsContext` in
+    `lib/loop-runtime.js`. Voice frame: name the bill exists in one line (cohort decides the
+    framing — first-product wants a number, vibe-virtuoso wants the inspect affordance,
+    domain-expert wants the privacy posture), point at `/ai-cost`, hand the decision back.
+    Never blocks. Override via devlog per IDEA-008.
+  - **`.gitignore`** updated to exclude `.boss/cost-log.jsonl` (local ledger; ship to a real
+    datastore when you have real users — the skill's review-cadence step says so).
+  - **L1-mvp manifest** now ships 8 skills + 4 loops + same 4 agents. `claude-append.md`
+    names the new skill and loop alongside the others.
+- **Pairs (not auto-invoked) with two existing mentors.** The handoff lines in `/ai-cost`:
+  *"`mentor-architect`, the cost shape says X — what architecture decisions does that imply?"*
+  and *"`mentor-business`, our cost-per-active-user is X; what should the pricing carry?"*
+  Cost discipline is the load-bearing connector between architecture (caching, batching,
+  cheaper-fallback) and business (unit economics, willingness-to-pay).
+- **The discipline reads as Husain-applied-to-spend.** *"Almost all AI cost problems are
+  visible in the ledger, and almost no one keeps one."* Liu cited for structured-outputs-as-
+  cost-lever; Mollick for cost-as-design-input. No new mentor or practitioner added — just the
+  application of existing discipline to a different artifact.
+- **What v0.25 does NOT do:** auto-instrument the user's code (judgment call: the founder
+  knows their call sites better; the skill *suggests* the wrapper, doesn't write it). Auto-
+  detect non-mainstream SDKs (LangChain wrappers, Replicate, Cohere, Bedrock). Ship a budget-
+  enforcement layer (this is a *nudge*, never a gate — IDEA-011 discipline applies).
+- **Conscience regression-clean.** Existing 43/43 conscience evals still pass — the new loop
+  lives in L1-mvp, the eval runner only loads L0-quickstart loops, so the new moment has no
+  way to fire against the existing eval fixtures. End-to-end tested in `/tmp`.
+
+## 0.24.0 — 2026-05-23
+
+- **Positioning pass — first non-feature release in BOSS's history.** The deliverable is *the
+  thinking* (positioning + README edit), not code. Per IDEA-012's revised roadmap, v0.24 was
+  explicitly the positioning pass, not a feature release. The Dunford exercise was first
+  recommended in the v0.15 advisory pass and deferred through 8 capability releases; finally
+  executed.
+- **Output:** [`docs/dossier/positioning-pass-001.md`](docs/dossier/positioning-pass-001.md) —
+  full Dunford exercise: target cohort (the founder using Cursor/Claude Code 3+ months with
+  2+ unfinished projects), 8 competitive alternatives, 8 unique attributes that survive
+  scrutiny, attributes mapped to value, 7 market-frame options, **8 candidate killer one-
+  sentences stranger-tested**, 8 cohort-tailored variants per persona, the trend layer
+  (*AI raised the speed limit; almost nothing raised the discipline limit*), 6 decisions to
+  act on, downstream consequences.
+- **Decisions on the record:**
+  - **Lead sentence (elevator):** *"BOSS is the just-in-time conscience for AI-native
+    founders. Pause it any time."* (13 words; killer for verbal intros.)
+  - **README opening sentence:** *"For founders building with AI — the thinking layer that
+    nudges when you're drifting and pauses on command. No growth-hacking pressure. Override-
+    friendly."* (22 words; killer for the README + landing-page hero.)
+  - **Category frame:** *"the thinking layer for AI-native founders"* — drops "incubator" as
+    the primary descriptor because it reads YC-shaped to strangers. "Incubator" stays valid
+    as a secondary descriptor.
+  - **Cohort-tailored variants:** 8 versions (one per persona archetype). Pattern: cohort-
+    naming first phrase + feature-that-lands-hardest second.
+  - **What BOSS doesn't compete on, named explicitly:** code generation (Lovable / v0 / Bolt).
+    BOSS is *complementary* to those — a founder could use Lovable to scaffold the app + BOSS
+    to scaffold the thinking about it.
+- **README updated** — the TL;DR replaced with the v0.24 positioning. Old: *"a just-in-time
+  incubator for AI-native projects."* New: the candidate #8 framing above.
+- **What this positioning is NOT** (recorded in the dossier):
+  - Not validated (synthetic-tested only until real founder reads + articulates back)
+  - Not permanent (evolves with the product)
+  - Not the only frame BOSS could use (picked AI-native-founder audience; the anti-VC
+    indie-hacker frame is a legitimate secondary positioning that could split into its own
+    surface later)
+- **Discipline applied to BOSS itself:** the Dunford exercise was on the record since v0.15.
+  Eight releases of capability shipped before it landed. The positioning pass finally
+  shipping is the discipline-on-the-discipline-tool move applied again — *not the most
+  exciting release; probably the highest-leverage one in months*.
+
+## 0.23.0 — 2026-05-23
+
+- **Conscience pause primitive — the discipline-on-the-discipline-tool move (IDEA-011
+  Phase 1).** Single-purpose release. Closes the canvas R&H #1 gap operationally: the founder
+  can now silence the conscience for a bounded session without having to edit settings.json
+  or rip out hooks. The four other items originally queued for v0.23 (Scale mode authoring,
+  moment #3 detector, PostToolUse hook plumbing, IDEA-010 Phase 4 `/design-prompt`) are
+  deferred to v0.24+ — see RESUME's restructured roadmap. **Discipline applied to BOSS
+  itself: ship only the one most-load-bearing thing in a release that *could* have been
+  bigger.**
+  - **`boss conscience pause [--for 8h | --until-resume] [--reason "..."]`** — records a
+    pause in `.boss/config.json`'s `conscience` block: `{ mode, since, expires, reason }`.
+    Default duration: 8h (a build session). `--until-resume` for indefinite pauses.
+  - **`boss conscience resume`** — explicit un-pause. Also happens automatically when the
+    recorded expiry passes (the hook auto-clears the state).
+  - **Hook reads pause state FIRST.** When `mode: paused` and `expires > now` → exit silent.
+    When `mode: paused` and `expires <= now` → call `clearPauseState` and continue normally.
+    The auto-resume IS the kindness; no special "your pause expired" signal (would be
+    performative noise; IDEA-011 explicitly chose silent auto-resume).
+  - **`boss status --conscience` surfaces pause state prominently** — at the TOP of the
+    output, marked with `⏸ PAUSED`, showing since/expires/reason. When expired but not yet
+    auto-cleared, shows `⏸ PAUSED (EXPIRED — will auto-resume on next prompt)`.
+  - **Help text updated** to include the new commands.
+  - **`readPauseState` + `clearPauseState`** added to `lib/loop-runtime.js` (template) —
+    the canonical version. BOSS's CLI imports from there so there's one source of truth.
+- **The architectural principle the pause demonstrates** (worth naming): **fractal-consistent
+  override discipline.** The same IDEA-008 grammar applied at two scales — per-loop overrides
+  in devlog (micro), and whole-conscience pause in config (macro). Same shape: deviation
+  conscious, recorded, never blocked, never forgotten (auto-resume is the kindness). Same
+  recipe, different scope. *Not novel as a pattern* (Focus modes are OS-level table stakes);
+  *worth claiming if anything* as the fractal-consistent application — Phase 3 externalization
+  may turn this into a publishable practice if BOSS gets used on > 1 project.
+- **End-to-end tested in /tmp — 5 flows verified:**
+  1. Baseline drifted state → hook fires (`caution low`)
+  2. `boss conscience pause --for 8h --reason ...` → state recorded
+  3. Hook fires after pause → silent ✓
+  4. `boss status --conscience` → shows ⏸ PAUSED prominently with since/expires/reason
+  5. `boss conscience resume` → hook fires again next prompt ✓
+  Plus expired-auto-clear: hook reads expired pause → clears it → emits signal normally next
+  prompt; config returns to `{ mode: 'active' }`.
+- **43/43 evals regression-clean.**
+- **Why a release for one feature?** Because it's the right discipline applied to BOSS itself.
+  The other four queued v0.23 items are real and queued for v0.24+. The pause primitive's
+  ratio of "leverage : code-size" is the highest of anything left in the roadmap (~50 lines
+  of code + tests; closes the canvas R&H #1 gap directly). MVP discipline: minimum experiment
+  that produces validated learning. Validated learning here: *did the pause primitive
+  actually make BOSS feel nimble?* Founder using BOSS to do an all-night build will tell us.
+
+## 0.22.0 — 2026-05-23
+
+- **V1 mode authored — `boss unlock v1` works for real.** The third major mode arrives. Same
+  playbook as MVP authoring (v0.14.0): manifest + template + claude-append + agents + skills +
+  loops. V1 is the *real shippable release* mode — design layer turns on, the second-tier
+  mentors arrive, data shape becomes a first-class decision, and a cross-FEAT sequencing
+  surface appears.
+  - **3 new builder agents:**
+    - `ui-designer` — token + visual authority. Reads `DESIGN_TOKENS.md` as truth; refuses raw
+      hex; three-layer architecture enforced (primitives → semantic → component). Cites Brad
+      Frost (Atomic Design), Nathan Curtis (layer-cake), Jina Anne (W3C DTCG), Diana Mounter
+      (Primer), Aarron Walter (emotional design). Holds the WCAG AA floor.
+    - `ux-designer` — flow + state + interaction authority. **5-state requirement** non-
+      optional (default / hover / active / disabled / empty + loading + error). Nielsen 10
+      heuristics; Krug clarity; Norman affordances; Luke Wroblewski forms; Christopher
+      Noessel for agentive patterns; Erika Hall for just-enough-research.
+    - `db-architect` — schema + data-shape authority. Schema before code, even solo. Cites
+      Codd, Date, Stonebraker, Kleppmann + AI-native data voices (Liu structured outputs,
+      Husain data quality, Huyen production). Flags AI-data failure modes (unstructured-LLM-
+      output in control flow; hallucinated-data pollution; eval-data-isn't-user-data).
+  - **4 template mentor copies** (promoted from BOSS-local v0.15.0 to scaffolded-project
+    templates — `{{PROJECT_NAME}}` / `{{MODE}}` placeholders): `mentor-business`,
+    `mentor-fundraising`, `mentor-pitch`, `mentor-talent`. Same source practitioners as the
+    BOSS-local versions; phrased to coach the founder of a generic scaffolded project. All
+    advisory; never binding legal/financial/medical. Default position for mentor-fundraising
+    + mentor-talent: *don't raise / don't hire yet, possibly never*.
+  - **3 new skills:**
+    - `/board` — cross-FEAT sequencing surface. Live read computed from FEAT frontmatter +
+      smoke/evals state + override entries. Flags: `--next`, `--blocked`, `--by-cohort`,
+      `--deferred`, `--evals`. Owned by `program-manager`.
+    - `/design-review` — before-code design review. Runs `ui-designer` (token + visual) + 
+      `ux-designer` (flows + 5 states) sequentially against the proposed UI. Reads tokens
+      file, style guide, canvas Promises cell. Outputs concrete diffs. Cohort-aware delivery.
+    - `/ux-check` — after-code UX review. Walks the *shipped* experience (not the spec),
+      checks states are real, runs accessibility heuristics, applies AI-specific UX where
+      relevant. Pairs with `/design-review`.
+  - **1 new loop:** `design-drift-loop` (V1-stage, runner_type: hook). The V1-stage counterpart
+    to MVP's `design-tokens-loop` — gates *whether tokens are still authoritative*, not just
+    *whether they exist*. **Subtle pattern worth naming:** this loop's exit predicate is the
+    *bad signal* (≥1 raw hex code in code, excluding the tokens file) — the loop is
+    "drift-emitting" when the bad signal is present. The IDEA-008 primitive supports this
+    without modification. Emits the `coherence` moment (introduced v0.21).
+  - **L2-v1 manifest** declares 7 agents + 3 skills + 1 loop; `boss sync` carries them via the
+    managed-file kinds (agent, skill, loop). `.boss` stamp tracks them. claude-append.md
+    reads as a clean V1-working-rules catalog.
+- **End-to-end tested in /tmp:** `boss new` → `boss unlock mvp` → `boss unlock v1` lands all
+  three modes' files cleanly. Final stamp: 14 agents + 15 skills + 6 loops + 1 hook + 3
+  installed layers. `boss status --conscience` shows all 6 loops in correct states on a fresh
+  project. 43/43 evals still pass (regression-clean).
+- **Deferred to v0.23:** moment #3 (capture — reusable value at breakpoint, needs LLM-as-judge
+  or heuristic detector — not predicate-based); PostToolUse hook for hardcoded-style detection
+  (new hook-type plumbing — its own concern); Scale mode authoring (mentor-humane template
+  promotion, PM org, code-health, product council); `/design-prompt` skill or fold into
+  `/design-review`.
+
+## 0.21.0 — 2026-05-23
+
+- **MVP discipline upgrades + IDEA-010 Phase 2 (design-tokens-loop) — all in one release.**
+  Three new skills, three new loops, one upgraded skill, moment #4 (restraint) lands skill-side.
+  Moment #3 (capture — reusable value at breakpoint) deferred to v0.22 — it needs a different
+  detector design (predicate evaluation doesn't fit "noticing this artifact is more general
+  than its loop").
+  - **`/spec` upgraded** (the smallest cut, highest leverage per the v1 playbook):
+    - Adds **validated-learning field** (Ries, *The Lean Startup*): "If this FEAT works
+      perfectly, what do we learn?" If the answer is "users like it" or "the feature works,"
+      don't build it. The MVP is the minimum experiment that produces validated learning, not
+      the minimum product to polish.
+    - Adds **evals field** (Husain): when a FEAT puts an LLM in control flow, the eval set
+      lives at `docs/evals/FEAT-NNN.yml`. Schema'd output (Liu) strongly recommended.
+    - **Moment #4 restraint check** (IDEA-008's collapsed-moments architecture): `/spec`
+      reads canvas-loop state before creating a FEAT spec; if canvas-loop isn't closed for
+      the active idea, the skill surfaces a Fitzpatrick-plain restraint nudge (cohort-aware
+      via v0.20's framing). Override grammar lives in devlog. Never blocks; always records.
+  - **`/evals` skill (new)** — Husain discipline as a first-class MVP skill paired with
+    `/smoke`. Smoke answers "is it alive"; evals answers "is it correct." Eval set first
+    (20 cases beats 0). Failure modes categorized by mode (Husain: failure modes are more
+    valuable than success modes). Structured outputs recommended (Liu: Pydantic-first).
+    Cites Husain + Liu + LLM-as-judge caveats.
+  - **`/pretotype` skill (new)** — Savoia's discipline as a first-class MVP skill. The
+    demand-test between `/canvas` and `/spec` — "make sure you're building the right It
+    before building It right." Six patterns named (fake door / WoZ / Mechanical Turk /
+    Pinocchio / YouTube test / impresario). The TRI metric (tangible / real-time /
+    imminent). Threshold-before-running (Ries pivot discipline). YODA (your-own-data >
+    anything).
+  - **`/design-tokens-init` skill (new)** — IDEA-010 Phase 2. Scaffolds the three-layer
+    token system at the first-UI-commit inflection. **Cohort-aware delivery** (v0.20's
+    framing): vibe-coder-newbie gets SHOWING; eng-builder gets OFFERING; vibe-virtuoso gets
+    OVERRIDE-FRIENDLY; first-product gets DEFINE-TERMS; indie-hacker gets RIGHT-SIZED;
+    returning-founder gets SKIP-THE-101; non-tech-founder + domain-expert get PLAIN-
+    LANGUAGE-COACH; unspecified gets neutral. The three-layer architecture (primitives →
+    semantic → component) is Curtis's layer-cake; the AI-tolerance argument is that
+    two-layer systems are fragile under AI generation (the field's consensus). Reads canvas
+    Promises cell to brand-anchor the primitives (prevents the brand-default problem from
+    IDEA-010).
+  - **Three new MVP-stage loops** on the v0.18 generic loop primitive:
+    - `spec-loop` (runner_type: skill) — encodes moment #4 restraint. Entry: canvas-loop
+      closed for some active idea. Exit: `FEAT-NNN-<slug>.md` exists. `/spec` is the
+      detector + runner.
+    - `pretotype-loop` (runner_type: skill) — structural by default (no drift_moment to
+      avoid over-firing). Records that demand-testing happened before significant build.
+      `/pretotype` is the runner.
+    - `design-tokens-loop` (runner_type: hook) — JIT, *only opens once UI starts
+      accumulating* (≥3 style declarations across src/). Drift moment: **`coherence`**
+      (new — system-vs-code drift; a flavor of caution specific to design-system mismatch).
+      Stack-agnostic regex catches common React/Vue/Svelte/Solid patterns; founders can
+      edit the spec for their stack.
+- **L1-mvp manifest** declares the new 3 skills + 3 loops; `boss sync` carries them; `.boss`
+  stamp tracks them. The `claude-append.md` reads as a clean catalog of what MVP offers.
+- **End-to-end tested in /tmp:** scaffold → unlock mvp → all 7 skills + 5 loops land + stamp
+  merges correctly + status --conscience shows the 5 loops in their right states (capture
+  open-structural, canvas/spec/pretotype/design-tokens all unopenable on a fresh project, no
+  spurious hook fires). 43/43 evals still pass (regression-clean).
+- **Deferred to v0.22:** moment #3 (capture — reusable value at breakpoint), V1 mode authoring
+  including the design-drift-loop (IDEA-010 Phase 3) + ui-designer + ux-designer agents +
+  /design-review + /ux-check + PostToolUse hook for hardcoded-style detection.
+
+## 0.20.0 — 2026-05-23
+
+- **The three design changes from v0.19's persona-reactions pass — landed.** Closes the loop on
+  the reactions: the personas flagged it, v0.20 shipped it. Moments #3 + #4 (the other v0.20
+  candidates per the published roadmap) deferred to v0.21 — they need their own design pass
+  (moment #3's "reusable value at breakpoint" detection is harder than predicate evaluation;
+  moment #4's "premature ceremony" needs `/spec`-skill-aware detection plus a spec-loop). The
+  three things shipping in v0.20 are the *persona-driven* design changes; the moment work is a
+  separate concern.
+  - **`boss status --conscience`** — the inspect affordance (asked-for by `eng-builder`,
+    `indie-hacker`, `vibe-virtuoso` personas in the v0.19 reactions doc). Loads
+    `docs/loops/*.md`, classifies each loop (open / closed / unopenable), shows what would
+    close the open ones (concrete predicate evidence: count/threshold, file matches, what's
+    missing), reads recent override entries from `docs/devlog.md` per IDEA-008's grammar,
+    shows the project's declared cohort. New `src/conscience.js` module formats the output;
+    the loop runtime is imported from the canonical Quickstart-template path so there's one
+    source of truth.
+  - **Cohort-aware conscience** — `.boss/config.json` carries an optional `cohort` field
+    (one of: vibe-coder-newbie | eng-builder | non-tech-founder | first-product |
+    vibe-virtuoso | indie-hacker | returning-founder | domain-expert | null). The
+    `loop-runtime.js`'s `composeContext` now reads the cohort and appends a **cohort framing
+    directive** to `additionalContext` — same signal, different voice. Each cohort gets a
+    distinct framing sentence: first-product needs *teaching*; returning-founder wants a
+    *harder cohort-aware question*; vibe-virtuoso gets *friction over praise*; indie-hacker
+    gets *plain Fitzpatrick language, not jargon*. The `/boss` spin-up skill now asks
+    one optional question to set the cohort; user can always edit `.boss/config.json` later.
+  - **Voice lineage decision: Fitzpatrick consistently.** The indie-hacker persona caught
+    that the prior conscience voice mixed Fitzpatrick ("who would you ask first") with
+    Maurya ("riskiest assumption" / "riskiest bet") in one breath. v0.20 picks Fitzpatrick-
+    plain ("what they'd want to learn"; "who they'd ask first") consistently. Updates the
+    `/triage` skill's exemplar text + the `composeContext` framing in `loop-runtime.js`. The
+    canvas itself still uses "riskiest assumption" (that's the canvas's frame, established);
+    the *conscience nudge* now speaks Fitzpatrick.
+- **Test coverage:** 43/43 evals still pass (regression-clean). End-to-end test in /tmp
+  verified: fresh project → capture-loop open (no signal, structural); 3 captures + no canvas →
+  canvas-loop drifts, hook fires with cohort framing in additionalContext; override in devlog →
+  appears in `boss status --conscience` output.
+- **What's NOT in v0.20** (now queued for v0.21): conscience moments #3 (capture — reusable
+  value) and #4 (restraint — premature ceremony). Moment #3 needs a different detector design
+  (predicate evaluation doesn't fit "noticing this artifact is generalizable"). Moment #4
+  needs a spec-loop authored AND skill-aware detection in `/spec`. Both warrant their own
+  release.
+
+## 0.19.0 — 2026-05-23
+
+- **Proto-personas layer + first reactions pass — the founder-experience eval channel.** 8
+  synthetic-founder agents now seated in BOSS's `.claude/agents/` with `persona-` prefix
+  (parallel to `mentor-`, parallel to the builder team). They REACT to BOSS features (not
+  advise, not mentor) so BOSS gets cheap pre-filter signal on how it lands across cohorts
+  *before* spending the expensive real-founder Mom Test call (which remains explicitly
+  overridden per advisory-pass #1; the override's re-open condition includes "persona
+  reactions surface a coherent product story" — this layer is how that signal arrives).
+  - The 8 personas: `vibe-coder-newbie` (no eng/startup background, picked up Cursor 3-6
+    months ago), `eng-builder` (10+ years eng, first-time founder, skeptical of magic),
+    `non-tech-founder` (deep domain expertise, can't code, AI is the bridge),
+    `first-product` (absolute beginner — to building, to vibe coding, to everything),
+    `vibe-virtuoso` (50+ shipped projects, zero sustained products, expert at idea-to-demo,
+    bad at company-building), `indie-hacker` (right-sized lens — Walling/Fried/Jarvis;
+    anti-VC; suspicious of venture-shaped language), `returning-founder` (has shipped before;
+    intolerant of 101 content; wants depth), `domain-expert` (medical/legal/financial;
+    stakes are real; humane lens applies hard).
+  - **`persona-reactions-loop`** authored in `docs/loops/` — captures the discipline (runner_
+    type: manual; uses the v0.18 loop primitive). Entry: persona agents exist. Exit: a
+    `docs/dossier/persona-reactions/<feature>.md` doc with structured reactions + synthesis +
+    design changes + real-founder questions the reactions sharpen.
+  - **First reactions pass complete** at
+    [`docs/dossier/persona-reactions/conscience-moment-1.md`](docs/dossier/persona-reactions/conscience-moment-1.md).
+    All 8 personas reacting to the conscience moment-1 firing scenario.
+  - **Three concrete design changes the reactions argue for** (ordered, with priorities):
+    1. **Cohort-aware conscience** — the model composing the conscience voice should know what
+       cohort the founder is in (`.boss/config.json` declares it, set by `/boss` during
+       scaffold). First-product needs *teaching*; returning-founder needs *a harder
+       question*; vibe-virtuoso needs *sharper architecture*. Same signal; different voice.
+    2. **Inspect affordance** — `boss status --conscience` or `boss conscience --explain` so
+       humans can see what loops are open / what would close them / what overrides exist.
+       Engineers + indie-hackers + vibe-virtuosos all asked for this. Plausibly v0.20
+       alongside moments #3/#4.
+    3. **Pick the lineage in the conscience voice** — current voice mixes Fitzpatrick
+       (talk-to-someone) and Maurya (riskiest-assumption) in one breath. Indie-hacker caught
+       this; the eval set didn't. Lean one direction consistently.
+  - **Three real-founder interview questions** the reactions sharpen for the eventual call:
+    (1) read-aloud comprehension test; (2) curious-vs-defensive-vs-confused for first-timer
+    cohort; (3) cohort-aware variant test for returning-founder cohort.
+  - **Two surprises from the reactions** that hadn't surfaced via any other discipline:
+    returning-founder wants a HARDER question, not a softer one (suggesting cohort-aware
+    direction); indie-hacker noticed the voice-fights-itself issue (Fitzpatrick vs Maurya
+    lineage mix) that 84 eval examples missed. **Argument for routine persona-eval passes on
+    every user-facing text.**
+- **Roadmap status:** v0.17 (builder team) + v0.18 (loop primitive) + v0.19 (personas) all
+  shipped this push. Next up: v0.20 (moments #3+#4 via the generic detector, leveraging the
+  v0.19 reactions for cohort-aware language) + v0.21 (MVP discipline upgrades).
+
+## 0.18.0 — 2026-05-22
+
+- **IDEA-008 → FEAT-001: generic loop runtime in Node; bash hook retired.** The biggest
+  architectural release since v0.8.0 (learning loop). The conscience hook used to be hand-coded
+  bash with a single hard-wired detector for moment-1. It's now a generic Node runtime that
+  reads `docs/loops/*.md` from the project, evaluates entry/exit predicates, and emits
+  structured signals — *any* loop drifting fires *its* moment, no per-moment code. Moments
+  #3 and #4 will be authored as loops (not detectors) in v0.20.
+  - **`conscience.js`** (replaces `conscience.sh`) — zero-dep Node, ~50 lines. Loads loops,
+    classifies state, composes signals, prints JSON. Fails silent — never blocks a prompt.
+  - **`lib/loop-runtime.js`** — the engine. Loads loops from `docs/loops/*.md`, evaluates a
+    closed predicate vocabulary (`exists`, `count_at_least`, `any_file_matches`) against the
+    live filesystem, classifies each loop as `unopenable | open | closed`, returns structured
+    signals only for hook-runner loops with a `drift_moment`. Loops without `drift_moment` are
+    *structural* (express dependencies but don't drift — caught the over-fires-on-fresh-project
+    bug live during the build; capture-loop is the canonical structural loop).
+  - **`lib/yaml.js`** — zero-dep YAML parser lifted from the eval runner so the same code parses
+    loop specs at hook time and eval examples at test time.
+  - **Two named loops authored** in the Quickstart template (`docs/loops/`):
+    - `capture-loop` — structural; expresses "at least one captured idea exists." Downstream
+      loops (canvas-loop) check this via their own entry predicates.
+    - `canvas-loop` — encodes moment-1's full logic *declaratively* via predicates. Entry: ≥3
+      dated capture-log entries across active (non-dropped) idea files. Exit: at least one
+      canvas tied to an active idea has a real (≥3-char alphanumeric) riskiest assumption.
+      Drift = open + not closed = moment-1 caution. The hand-coded bash logic is now expressed
+      in YAML.
+  - **`manifest.json` gains a `loops` array.** `src/scaffold.js` + `src/sync.js` handle loops
+    as a new managed-file kind (`kind: loop`, `rel: docs/loops/<name>.md`). Hook-library files
+    (`lib/*.js`) auto-discovered from the template and synced alongside the hook script. The
+    `.boss` stamp tracks `loops` (alongside agents/skills/hooks).
+  - **`runner_type` field on loop specs** — resolves the moment-2 shape question from v0.16's
+    meta-learnings. Today only `hook` is honored by the conscience runtime; `skill`, `manual`,
+    `external` will be honored by future runners (skill-prompt eval, manual review, external
+    detector — e.g. CI).
+  - **Settings migration (bash → node).** Existing projects pinned at <= 0.13.0 have
+    `bash …conscience.sh` in their settings.json. The merge logic now applies hook-command
+    *migrations* before the additive merge: it drops stale bash entries before adding the new
+    node entry. Tested with a synthetic pre-0.18 project: bash entry dropped, node entry added,
+    user's permissions preserved, idempotent on re-sync.
+  - **Eval set unchanged; 43/43 still pass against the generic runtime** (regression-coverage
+    in place). The runner now invokes `node conscience.js`, materializes `docs/loops/` into
+    the test project before each example (so the runtime has loops to load), and the existing
+    examples test the *generic detector* end-to-end. One bug found during the build (POSIX
+    regex char classes — `[[:space:]]` doesn't work in JS) fixed by switching loop specs to
+    JS-native regex (`\s+`, `[a-zA-Z0-9]{3,}`). Eval-first discipline catching itself.
+  - **BOSS dogfoods:** `docs/loops/capture-loop.md` + `docs/loops/canvas-loop.md` now live in
+    BOSS itself (alongside the existing `docs/loops/eval.md`). BOSS-the-project runs the same
+    runtime against its own state.
+  - **Verdict on the primitive (from v0.16's meta-learnings): confirmed under contact with
+    real Node implementation.** The four-field shape (entry / purpose / exit / drift) held up.
+    Predicate vocabulary covered everything moment-1 needed declaratively. The structural-loop
+    pattern (no `drift_moment`) handles dependencies-only cases cleanly. Net code change:
+    `conscience.sh` (40 lines bash) replaced by `conscience.js` + `lib/` (~250 lines Node) —
+    bigger surface, but moments #3+ will be loop specs (~30 lines YAML each), not new detectors.
+
+## 0.17.0 — 2026-05-22
+
+- **Builder team seated alongside the mentor board.** Three new builder agents in BOSS's own
+  `.claude/agents/` — they make BOSS *feel right*, parallel to the mentor board which makes
+  BOSS *be right*. Builders, not mentors: they propose concrete diffs, not advice. Cover the
+  interaction-design layer that wasn't covered by either the mentor board or the existing builder
+  agents (pm, coder-generalist, tester, program-manager).
+  - **`designer`** — owns the UX of the entire BOSS interaction experience. Not "the visual
+    designer" (BOSS is a CLI + Claude Code experience, not a webapp). Owns: when BOSS speaks vs
+    stays quiet, what a skill *feels* like when run, the rhythm of mode unlocks, the surprise
+    vs predictability of conscience moments, what the founder is being asked to *do* vs read at
+    every step. Sources: Norman, Krug, Nielsen, Spool, Wroblewski, Walter + AI-specific UX
+    heuristics (options-not-truth; visible confidence; deliberate failure states).
+  - **`voice-keeper`** — guards what BOSS *sounds like*. Reviews skill text, agent system
+    prompts, hook signal language, README, CHANGELOG, error messages. Catches performed warmth,
+    scolding tone, voice-mode bleed, framework-jargon leaking into user-facing text, assumed
+    knowledge, hedging. Proposes concrete edits side-by-side. Inward-facing language guardian.
+    Sources: `boss-voice` memory (canonical spec), Strunk & White, Raskin + Neumeier (pitch =
+    product voice), Godin (write to one person).
+  - **`prompt-coach`** — helps the founder write better prompts (to BOSS, to Claude, to AI in
+    general) and teaches the craft over time. Outward-facing counterpart to voice-keeper.
+    Builds a per-founder pattern library in `docs/dossier/founder-prompt-patterns.md`. Catches
+    vague / multi-prompt / missing-constraint / missing-output-format / leading-question /
+    missing-context / wrong-role-assignment / stop-word-missing failure modes. Sources:
+    Karpathy (think in distributions), Mollick (AI-as-different-roles), Willison (prompt-as-
+    code), Liu (Pydantic-first), Husain (look at the output), Fitzpatrick (Mom Test discipline
+    applied to interview prompts).
+- **Advisory-pass #1 (real-founder Mom Test calls) explicitly overridden — first real use of
+  IDEA-008's override grammar.** Ajesh's call: at zero users + product still defining its shape,
+  expensive real-founder calls are premature; cheap synthetic signal from proto-personas (v0.19
+  work) is the right move now. Override recorded in `docs/dossier/boss-advisory-pass-001.md`
+  with explicit re-open conditions (persona reactions surface a coherent product story, OR a
+  non-Ajesh user starts using BOSS in earnest, OR the eval set catches something only real-
+  founder feedback could surface). The recommendation isn't deleted — it's deferred under the
+  IDEA-008 contract: deviation made conscious, recorded, re-openable.
+- **Roadmap published** (in RESUME): v0.17 builder team → v0.18 generic loop primitive (IDEA-008
+  to FEAT) → v0.19 proto-personas as named loops → v0.20+ moments #3/#4 + MVP discipline
+  upgrades + V1/Scale mode authoring + IDEA-003 finish + externalization + backlog. Ten releases
+  sequenced for build-on-build; the discipline rails (evals, structured output, loops, personas)
+  make each one buildable in a focused session.
+
+## 0.16.0 — 2026-05-22
+
+- **The eval-loop closed — conscience now has evals + structured output (proves IDEA-008's
+  primitive on its first real run).** Two ladders climbed at once: produced the conscience eval
+  set (the artifact the advisory pass / playbook said BOSS most needed) AND validated the loop
+  primitive from IDEA-008 by running ONE loop end-to-end. Both succeeded.
+  - **84 labeled eval examples** (43 moment-1 caution + 41 moment-2 Done!) in
+    `docs/architecture/conscience-evals/{moment-1-caution,moment-2-done}.yml`. Should-fire +
+    categorized should-NOT-fire entries (failure_mode taxonomy per Husain's discipline:
+    `over-fires-on-fresh-project`, `fires-mid-other-work`, `repeats-itself`, `shame-toned`,
+    `false-positive-canvas-exists`, `false-positive-not-drift`, `acknowledged-already`,
+    `fires-too-early`, `performed-warmth`, `removes-agency`, `riskiest-assumption-stale`,
+    `triggered-by-trivial-change`). Each example has structured `project_state` (the synthetic
+    docs/ideas/ tree the runner builds in a temp dir) + `expected_detection`.
+  - **Zero-dep Node runner** at `docs/architecture/conscience-evals/runner.js` — includes a
+    minimal YAML parser for the subset our eval files use (so the data stays human-readable
+    without breaking the zero-dep rule). Constructs synthetic state, invokes the actual hook,
+    parses output, asserts. Reports per-example + a categorized summary table.
+  - **Hook refactored to structured output** (Liu's discipline): now ships
+    `{moment, confidence, evidence: {capture_count, canvases_with_filled_assumption,
+    active_idea_count}, suppress_if}` in addition to the `additionalContext` string. Model still
+    composes the voice; hook ships a schema.
+  - **3 real bugs caught and fixed by the eval set itself** (Husain's "look at your data"
+    discipline in action):
+    1. Single-char placeholders like `?` slipped through the riskiest-assumption regex.
+       Tightened to require ≥3 alphanumeric chars (rejects `?`, `??`, `_TBD_`, empty, etc.).
+    2. Hook counted captures across *all* ideas — including `status: dropped`. Drift signal
+       from already-walked-away ideas is meaningless. Hook now filters active vs. dropped.
+    3. Filled canvases on dropped ideas were stopping the "validated" check. Same fix:
+       active-status filtering on canvas checks too.
+  - **Runner results: 43/43 pass on every runnable case. 41 skipped as categorized future-work**
+    (moment-2 lives in `/canvas` skill prompt not a hook; suppress_if cases need session-state /
+    devlog awareness; signal-text violations need a separate runner — all explicitly tracked).
+  - **IDEA-008 primitive verdict — ready to promote to FEAT.** The four-field shape (entry,
+    purpose, exit, drift) held up. The predicate vocabulary (`exists`, `contains`,
+    `count_at_least`, `recorded_at`) survived contact with reality. Multi-part exit artifacts
+    work. Skip-with-reason is the right runner pattern. ~half of examples test future features
+    (suppress_if + devlog + skill-based detection), which is the right ratio — the eval set
+    is forward-looking, not just current-implementation snapshot. One real shape-question
+    surfaced (moment-2 isn't hook-detected — argues for a `runner_type` field on loop specs).
+  - First loop authored: `docs/loops/eval.md` — the eval-loop spec using the four-field primitive
+    with predicates in YAML frontmatter. The template for every future loop.
+
+## 0.15.0 — 2026-05-22
+
+- **BOSS now has its full mentor board seated, and the board has had its first session on BOSS
+  itself.** Step back before the next build axis: identify and instantiate all the experts we'd
+  want consulting on how to build an AI-native MVP right, then have them actually review BOSS.
+  - **8 mentor agents live in this repo's `.claude/agents/`** (BOSS-local; the project-as-its-own-
+    founder): `mentor-venture` (cornerstone), `mentor-architect`, `mentor-gtm` (the three that
+    already existed in stage templates, now BOSS-tuned), plus 5 new: `mentor-business`,
+    `mentor-fundraising`, `mentor-pitch`, `mentor-talent`, `mentor-humane`. Each cites the
+    practitioners from `docs/mentor-practitioners.md` it draws on — no agent impersonates a
+    person; mentors cite named practices (per the encoding decision in `docs/MENTORS.md`).
+  - **`mentor-humane` carries explicit override authority** when a humane concern is on the table
+    — the conscience's conscience. Seated from day one despite the Scale-level slot in the
+    roster, because BOSS itself is the special case (it has to *be* humane in its construction,
+    not just preach it).
+  - **`mentor-architect` retuned for the AI-native MVP era** (both BOSS-local and the MVP template
+    in `stages/L1-mvp/template/.claude/agents/`). Leads with AI as the modality; classical-stack
+    choices are supporting cast. Names the load-bearing AI questions: surface, eval strategy,
+    prompt vs. fine-tune vs. RAG, structured outputs, human-in-loop boundaries, cost/latency
+    budgets, fallback. Cites Karpathy, Willison, Husain (evals — load-bearing), Liu (structured
+    outputs — load-bearing), Mollick, Huyen.
+  - **First advisory pass captured at `docs/dossier/boss-advisory-pass-001.md`.** Honest, not
+    flattering. Each mentor's read on BOSS as-of v0.14.0, citing their practitioners. Five
+    cross-cutting themes converged: (1) pause "more features" to earn founder contact; (2) evals
+    are the next architecture investment (moments #3/#4 should be eval-set-first); (3)
+    right-sized is the default shape (calm-company / OSS / patronage, not venture); (4) interior
+    story rich, exterior story missing (strangers can't read the README); (5) the conscience is
+    the moat *and* the most under-validated thing — plug this gap first.
+  - **Next moves (re-ordered by the pass — supersedes the prior queue):** conscience-evals doc +
+    structured hook output, *then* moments #3/#4; 5 real-founder Mom-Test interviews; Dunford
+    positioning exercise + strangers-can-read-it README; humane upgrades to the conscience spec
+    (cumulative-pressure check, BOSS.DK exemplar lines, cross-link humane practitioners into
+    architect's lens); name the right-sized shape on the canvas.
+
+## 0.14.0 — 2026-05-22
+
+- **MVP mode (L1-mvp) is authored — `boss unlock mvp` works for real (closes IDEA-002).** Until now
+  the L1 stage was a placeholder README; unlocking MVP errored out as "not authored yet." This release
+  extracts this repo's own MVP practice UP into a real `stages/L1-mvp/{manifest.json,template/}`:
+  - **Skills:** `/spec` (promote `IDEA-NNN` → `FEAT-NNN` with goal, acceptance criteria, smoke check,
+    out-of-scope), `/smoke` (stack-configured build-health gate; saves the command to `.boss/smoke.json`
+    on first use), `/log` (append-only `docs/devlog.md` — newest at the top, dated, FEAT-tagged),
+    `/close` (session-end ritual that rewrites `docs/RESUME.md` and runs `/log`).
+  - **Builder agents:** `tester` (owns `/smoke` + FEAT acceptance — surfaces, doesn't fix),
+    `program-manager` (the *when* — sequences FEATs, names blocks, distinct from `pm`'s *what*).
+  - **Mentor agents (advisory, never code):** `mentor-architect` (load-bearing decisions: data shape,
+    boundaries, what to defer — the calibration against over-architecting an MVP) and `mentor-gtm`
+    (first 100 users, channels, message — earned-when-needed, humane before viable).
+  - **Additive CLAUDE.md** via `claude-append.md` — the mechanism shipped in v0.8.0 finally has its
+    first real consumer. MVP's working rules (spec → smoke → log → close loop, conscience still runs)
+    fold into the existing CLAUDE.md under the `boss:L1-mvp` marker; never overwrites Quickstart's rules.
+  - **`boss sync` carries it for free** — the manifest's agents/skills are picked up by `managedFiles`
+    in `src/sync.js`, so projects pinned at older versions get the MVP files via `/boss-sync` once they
+    unlock the layer. (Hooks list is empty in this manifest — moments #3/#4 remain TBD.)
+  - Tested in `/tmp`: scaffolded a Quickstart project, `boss unlock mvp` added 4 skills + 4 agents,
+    appended the MVP working-rules block to CLAUDE.md, updated the stamp (`L0-quickstart → L1-mvp`,
+    merged agents/skills), kept Quickstart files untouched, re-running unlock no-ops cleanly, `boss
+    sync` recognizes everything as up-to-date.
+
+## 0.13.0 — 2026-05-21
+
+- **`boss sync` now carries hooks + settings (closes the conscience's reach gap).** Until now the
+  conscience hook (v0.12.0) reached *new* projects only — `boss sync` carried skills/agents but not
+  hooks or `settings.json`, so existing projects (e.g. `betabeta`) couldn't get it. Fixed:
+  - **Hook scripts sync like any managed file** — `manifest.hooks` → `.claude/hooks/<name>.sh`, shown as
+    `new`/`changed`/`ok` in the preview, written on `--apply`.
+  - **Hook registrations merge into `.claude/settings.json` additively** — `boss sync` adds the
+    `UserPromptSubmit` (etc.) entries BOSS ships, **matched by command so it's idempotent**, and
+    **preserves the user's permissions and their own hooks.** It's the one user-editable file sync
+    touches, and only the `hooks` block. (`computeSettingsMerge` in `src/sync.js`.)
+  - The `.boss` stamp now tracks `hooks` (alongside agents/skills); `boss new`/`unlock` record them.
+  - `/boss-sync` skill narrates the new behavior. Tested in `/tmp`: an "old" project (no hook,
+    permissions-only settings, pinned 0.10.0) → sync adds the hook + merges settings (permissions
+    preserved) + bumps the pin; idempotent on re-run.
+  - **Existing projects can now `boss sync --apply` (or `/boss-sync`) to receive the conscience.**
+
+## 0.12.0 — 2026-05-21
+
+- **The conscience can now speak *unprompted* (spike → shipped).** Until now both moments only fired
+  when you ran a skill (`/triage`, `/canvas`). A new **`UserPromptSubmit` hook** lets moment #1
+  ("what does this prove?") surface on its own:
+  - `.claude/hooks/conscience.sh` — detection only: if ≥3 dated capture-log entries exist across
+    `docs/ideas/` and no canvas has a *filled* riskiest assumption (capturing-lots / validating-nothing),
+    it returns `additionalContext` — a **signal**, not canned copy. Claude keeps the voice and the
+    judgment: it decides whether the moment fits, says it once in BOSS's register, or stays silent.
+  - Registered in the template `.claude/settings.json`; invoked via `bash …` so it needs no execute bit.
+    `manifest.json` hooks: `["conscience"]`.
+  - Confirms the architecture for the remaining moments: **hook = detection, model = tact + voice.**
+  - _Caveats:_ reaches **new** projects only (`boss sync` doesn't carry `settings.json`/hooks yet); the
+    *feel* (wise vs. naggy) still needs live validation, like moment #1 in `betabeta`.
+
+## 0.11.0 — 2026-05-21
+
+- **The conscience — second moment: "Done!" (`/canvas` graduation).** The *affirming* register,
+  counterpart to moment #1's caution. When the canvas holds up (most cells real + riskiest assumption
+  has a validation plan), `/canvas` no longer just offers `boss unlock mvp` — it marks the threshold in
+  two beats:
+  - **Arrival** — names what became real (started with a hunch → now a specific person, real tension,
+    sharp promise, a testable riskiest bet). Said plainly, let to land. Acknowledgment, not praise.
+  - **Next doorway** — points at `boss unlock mvp` (build tools + next mentors) without rushing; the
+    canvas keeps.
+  - A threshold, not a finish line; never forced — the celebration is for when it's genuinely earned.
+  - Completes the conscience's *two registers* (caution + completion) in Quickstart.
+
+## 0.10.0 — 2026-05-21
+
+- **The conscience — first moment lands (`/triage` validation check).** BOSS starts behaving like the
+  *build's conscience*, not just a set of skills you invoke. The first of four conscience moments —
+  **"what does this prove?"** — is now baked into `/triage`:
+  - **Fires when** the active idea has ≥3 capture-log entries and no canvas with a filled riskiest
+    assumption (the "capturing lots, validating nothing" drift) — and *only* then.
+  - **Says one spare line** in BOSS's voice (the seasoned hand): names the drift, asks what would make
+    it real, points at `/canvas`, hands the decision back. Never blocks a capture, never nags.
+  - Turns the validation thinking that already lived in `/canvas` + `mentor-venture` (invoke-only) into
+    a *moment that fires* in the flow where drift actually happens.
+  - Template `CLAUDE.md` names the conscience in the Quickstart arc.
+  - Existing projects pick it up via `boss sync` / `/boss-sync`.
+
+## 0.9.0 — 2026-05-21
+
+- **Mentor layer — structure + cornerstone (IDEA-003).** BOSS's second agent class lands.
+  - **`docs/MENTORS.md`** — the design: two classes (builders make the app, `mentor-*` coach the
+    founder), the roster + JIT-per-mode mapping (venture → architect/GTM → fundraising/pitch/talent/
+    business → humane), the founder **dossier** (canvas → proposal → architecture brief → pitch →
+    hiring plan → data room), and the hard line (no binding legal/financial advice; humane before viable).
+  - **`mentor-venture` agent** seeded into Quickstart (`library/agents/` + template + manifest). The
+    cornerstone mentor: pressure-tests whether an idea is worth it, names the riskiest assumption,
+    points at the next real step, owns the canvas conversation. Advisory only — never writes code/specs.
+  - Existing projects pull `mentor-venture` + the new skills via `boss sync` / `/boss-sync`.
+  - _Still open:_ encoding real practitioners' best-practices UP into `practices/` + `memory-seed/`
+    (awaiting the list); authoring the rest of the roster as their modes get built.
+
+## 0.8.0 — 2026-05-21
+
+- **The learning loop (IDEA-001).** PRINCIPLES #1 made operational, in both directions:
+  - **`boss sync [--apply]`** — brings a project's BOSS-managed skills/agents (across all its
+    installed modes) up to the current version. Previews as `new` / `changed (N lines)` / up-to-date,
+    reconciles stale mode labels (an old `L0-sketch` pin → `L0-quickstart`), and on `--apply` writes
+    the files + bumps the project's `.boss` pin. Syncs BOSS-managed skills/agents only; user-editable
+    files (`CLAUDE.md`, `settings.json`) are left for hand-merge.
+  - **`boss learn <path> --as <category>`** — promotes a proven pattern UP into
+    `library/<category>/` (`agents|skills|hooks|practices|memory-seed`), bumps `VERSION` +
+    `package.json` in sync, and prepends a CHANGELOG stub. Finds the BOSS **source** repo via the
+    registry's self-hosted entry (or `$BOSS_SRC`), so it works when `boss` runs from a global install.
+  - **`/boss-sync` + `/boss-learn` skills** — the judgment layer. `/boss-learn` is a two-destination
+    **router** (UP = BOSS superset practice; DOWN = harden into the app's own core), never one-way.
+    `/boss-sync` narrates the diff from the CHANGELOG and flags local edits before overwriting.
+- **`claude-append.md` support in `boss unlock`.** A mode template can carry a `claude-append.md`
+  whose contents are *appended* to the project's CLAUDE.md under an idempotent marker — additive
+  working rules, never an overwrite. (Needed by the MVP mode next.)
+
+## 0.7.0 — 2026-05-21
+
+- **Package / dogfood separation.** Clean boundary between the shippable package and BOSS's own
+  incubation layer:
+  - `package.json` `files` allowlist — only `bin/ src/ stages/ library/ VERSION PRINCIPLES.md
+    registry/CHANGELOG.md` (+ README/LICENSE) ship. `docs/`, `.boss/`, root `CLAUDE.md`, and the
+    local registry are never published. (`npm run pack:preview` to verify.)
+  - **Registry moved out of the repo** to `~/.boss/registry.json` — machine-local state (your
+    project list + absolute paths) no longer lives in (or leaks into) the package/repo.
+  - `VERSION` and `package.json` version synced to 0.7.0.
+
+## 0.6.0 — 2026-05-21
+
+- **BOSS now dogfoods itself.** BOSS is its own first registered project (`.boss/` stamp,
+  mode MVP, self-hosted) — retrofitted ahead of the MVP-mode template, which will be *extracted UP*
+  from this repo's working practice (Principle 1).
+- Added BOSS's own dogfooded docs: root `CLAUDE.md`, `docs/IDS.md`, `docs/ideas/` (IDEA-001 learning
+  loop, IDEA-002 MVP mode, IDEA-003 mentor layer), `docs/ideas/CANVAS.md` (BOSS's own Humane
+  Product Canvas), and `docs/RESUME.md` (multi-session continuity).
+- Recorded the **mentor layer** vision: two agent classes — builders (make the app) and mentors
+  (coach the founder); mentors accumulate a founder dossier toward funding/hiring.
+
+## 0.5.0 — 2026-05-21
+
+- **PRINCIPLES.md** — BOSS's six operating principles. #1: always scaffolding, but pause to sort
+  patterns UP (BOSS superset practice) or DOWN (app core); `/boss-learn` becomes a two-way router.
+  #3: nothing valuable gets locked into code (style → tokens, prototypes reuse the same system).
+- **Design-system practice** (`library/practices/design-system.md`) — generalized from dhun:
+  tokens as single source of truth, central style utils, 5-state rule, prototype reuse, JIT
+  enforcement (turns on at V1; seed tokens the moment real UI appears).
+- V1 mode stub fleshed out with the design layer + enforcement timing.
+
+## 0.4.0 — 2026-05-21
+
+- **Quickstart becomes a tiny incubator:** capture → keep adding → canvas → unlock MVP.
+- `/triage` rewritten as **living idea capture** — one evolving doc per idea (sharpening
+  "current shape" + append-only capture log). Re-run anytime to keep adding.
+- New `/canvas` skill: a **humane business pressure-test** — Ajesh Shah's Humane Product Canvas
+  (Human Foundation / Product Expression / Stewardship) as the spine, with Lean + Lenny-style
+  prompts folded into each cell, plus the incubation heartbeat (riskiest assumption + one
+  experiment this week). Acts as the Quickstart→MVP graduation gate.
+- North star recorded: BOSS is a just-in-time startup incubator — the right support shows up per mode.
+
+## 0.3.0 — 2026-05-20
+
+- Stages renamed to **modes** (the user's vocabulary): Quickstart → MVP → V1 → Scale
+  (folder ids `L0-quickstart` / `L1-mvp` / `L2-v1` / `L3-scale`).
+- `boss unlock` accepts mode names, levels, or full ids (`mvp`, `L1`, `L1-mvp`).
+- `boss new` / `status` / `list` display the mode name; `.boss/manifest.json` records `mode`.
+- _Migration note:_ projects created on ≤0.2.0 carry the old `L0-sketch` stage label in their
+  stamp; cosmetic only — `unlock`/`status` still work. A future `/boss-sync` will reconcile labels.
+
+## 0.2.0 — 2026-05-20
+
+- `/boss` spin-up skill (L0): reads a PRD/idea → shapes via pm lens → captures IDEA-001 →
+  recommends stack + stage → optionally creates a **private** GitHub repo with a license.
+- Repo-creation defaults (in `.boss/config.json`): `github: ask`, `visibility: private`,
+  `license: proprietary` (All Rights Reserved — keeps paid + OSS options open; relicense later).
+- Auto-sets repo-local GitHub noreply email to avoid the GH007 email-privacy push block.
+
+## 0.1.0 — 2026-05-20
+
+- Walking skeleton: `boss new` / `unlock` / `status` / `list` / `version`.
+- L0 · Quickstart stage authored (CLAUDE.md, pm + coder-generalist agents, /triage, ideas pool, IDS).
+- Registry + `.boss/` project stamp + version-pinning.
+- L1–L3 stages stubbed (manifests/templates not yet authored).