v0.6.0

The first stable release under the new name (0.6.0-rc.1 claimed tokenroom on npm
2026-07-01). Gated on — and released with — the batched ADR-9 wording eval for every
post-0.3 wording item (ADR-19/20/22/23/24): PASSED, all equipped assertions green on
the Sonnet tier with clean naive-harm baselines for the ADR-24 items
(eval/v3-wording/results/2026-07-02-batched-post-0.3-wording.md).

Fixed (since 0.6.0-rc.1)

• Account-switch banner no longer false-fires on a same-account window rollover, and
  the pair advisor will never recommend switching to the profile you are already on
  (ADR-24 hardening).
• Hooks never mislead: every quota-consuming hook honors the per-account show-gate
  (ADR-21), and a malformed resume.json suppresses the readiness stamp instead of
  garbling it (ADR-5).
• Eval harness: eval/v3-wording/ gained the seven batched post-0.3 scenarios
  (S-R/S-G/S-B/S-Q/S-W/S-E/S-K) + rubric.

Added

• Multi-account profiles + smart switch advice (ADR-24). For users toggling two
  subscription accounts via /login:
  • Instant switch detection — the payload is ground truth: a statusline render whose
    account key differs from the session's mapping remaps in that same tap invocation
    (payload wins, always), logs an account_switch event, and the next stamp announces
    the switch once with the NEW account's numbers. Fixes the field capture 2026-07-01
    where a session stamped "0% left" for ~20 minutes after switching to an account at 98%.
  • Echo honesty — the tap tracks when window VALUES last moved
    (values_changed_at); a critical figure frozen >5 min while a sibling account has
    values-newer data is stamped as a possible pre-switch echo ("figures refresh on the
    next completed turn"), never asserted as fresh.
  • Named profiles are identity (ADR-21's phase keys give isolation, not identity):
    tokenroom account label <name> / list / fold <key> <name> /
    config-dir <name> <path>, stored in ~/.tokenroom/profiles.json. Heuristic folding
    only ever HINTS (doctor + account list); zero profiles → behavior unchanged.
  • Pair-aware descent — low active window + a fresh other profile → stamps and
    mid-turn bands say "finish this unit at full speed, then switch (/login or
    tokenroom switch) for zero downtime; defer only if BOTH profiles are thin"; the 1%
    floor becomes land-and-switch. Healthy active → HUD-only terse alt '<label>' ≈X%.
  • tokenroom switch (profile decision table + recommended move) and
    tokenroom run [--profile X] (launch claude under the best/named profile's
    config dir via the official CLAUDE_CONFIG_DIR env var — launch-time selection only;
    auth files are never read or written, nothing is swapped mid-session).
  • New stamp/advice wording joins the batched ADR-9 eval before the npm release.

Fixed

• MCP tools no longer report whichever account rendered last (ADR-21 gap). The server
  now resolves quota via the tap's session map: one recently-active account → that
  account's state; two or more active in the last 10 minutes → quota is withheld with an
  explicit attribution: "ambiguous — quota withheld …" on
  resource_state/estimate_remaining/fit_check/plan_resume instead of ever
  returning the wrong account's numbers.

Changed

• Renamed: headroom → tokenroom (ADR-23). The practical owner of the name "headroom"
  is headroomlabs-ai/headroom (55k★, whose CLI binary is literally headroom — a direct
  bin conflict), with a commercial extraheadroom.com in the same niche; tokenroom was
  verified free on npm 2026-07-01. Everything user-visible renames with it: package
  tokenroom, bin tokenroom, state dir ~/.tokenroom (installer COPIES ~/.headroom
  across once, never moves/deletes it), env TOKENROOM_DIR/TOKENROOM_DISABLE (clean
  break), [tokenroom] stamp prefix (mechanical branding; ADR-9 eval batched to the
  harden round), MCP server tokenroom, skill tokenroom, repo
  github.com/tyejcoleman/tokenroom. Install now REPLACES any pre-rename headroom
  statusline/hooks/skill/MCP/CLAUDE.md-block artifacts in place — never duplicates them —
  and uninstall removes both namings; doctor hints while a stale ~/.headroom remains.
  Historical records (past CHANGELOG entries, prior ADR bodies, eval results/fixtures)
  keep the old name.

Removed

• ARM mode (armed resume, the autonomous headless executor) — removed entirely
  (ADR-22, supersedes ADR-16). src/arm.mjs, the launchd plist machinery, headless
  claude -p invocation, headroom resume --arm/--disarm, the resume-run entry
  point, the auto_arm config flag, and their tests are gone. The 2026-06-15 platform
  change moved programmatic use (claude -p, SDK) to a separate monthly API-priced
  credit pool, so an armed resume no longer salvages expiring subscription quota — it
  spends new metered credits. Autonomous continuation moves to the separate Conductor
  package on official in-session surfaces (Stop hook + scheduled wakeups + cron
  routines), carrying ADR-16's consent principle with it. The awareness half stays
  untouched: plan_resume, the resume-plan file, the HUD reset countdown /
  ✓ deferred work ready, readiness stamps, and headroom resume [--clear].

v0.5.1

Fixed

• Weekly "HOT pace" no longer false-fires right after the reset. enrichWeekly
  flagged hot whenever the extrapolated pace exceeded 1.15× sustainable, even when only
  a sliver of the week had elapsed — so a short burst in the first hours (e.g. 4% used in
  ~4h) projected "you'll exhaust the week" with 96% left, which isn't actionable. HOT now
  additionally requires that a material share of the weekly budget is actually used
  (seven_day.used_pct >= 15, ≈ one day's sustainable allowance); below that the stamp
  shows the calm "cruising" form. Deliberately a usage floor (not an elapsed-time floor),
  so a genuine heavy early burn still surfaces.

v0.5.0

Changed

• Aggressive descent on the rate-limit window (ADR-19). The agent now works at FULL
  SPEED until 5% remains (was: "descend, small atomic steps" from 10%), is told to be
  velocity-mindful but keeps working from 5% down to a 1% floor, and only does
  finishing-moves at ≤1% (was: "start nothing new" at 5%). The 1–5% band keeps a stranding
  guard (prefer small divisible steps, checkpoint often, defer huge/indivisible new tasks),
  and the velocity-aware overrides still keep full speed when the window resets before it
  would exhaust. Adds a 1-band to every governor profile so the floor message fires.

Added

• Multi-session burn awareness (ADR-20). The stamp now discloses the combined burn
  rate across all sessions (≈tok/min) and how many are actively burning, and flags an
  anomalous burner — a session burning ≥3× the median of the others — naming whether
  it's this session (ease off) or another (the shared window can drop fast; re-check
  often). Flow samples are tagged by session_id to attribute per-session burn.

v0.4.2

Changed

• Two resources, two opposite postures — made unmistakable. Context is a burn-through
  resource; quota (rate-limit) is a wary, paced one. The mid-turn "context getting low"
  nudge now leads with "BURN IT, don't conserve it" and states outright that low context
  is never a reason to slow, stop, hand back control, wrap up, or get cautious — the only
  thing it asks is that you keep the handoff current, then keep working at full speed until
  auto-compaction fires and refreshes you. It closes by contrasting the two budgets so the
  agent never applies quota-caution to context (or context-fearlessness to quota). The
  intelligent, reset-aware quota guidance (descend/defer unless the window resets before
  you'd run dry) is unchanged — that resource stays wary.
• Default context nudge held to the core (~4% left). ctx_bands default lowered from
  [8, 4] to [4], so the handoff nudge fires once near the ceiling instead of at 8% —
  the agent uses nearly all its context before prepping a handoff. The token-floored
  "super close to auto-compaction" message remains the final safety net; powersave keeps
  its earlier 10% heads-up for thrift.

v0.4.1

Changed

• Power through to auto-compaction — stop getting "scared" near the context ceiling.
  Field report: agents were stopping/handing back control as context filled, which prevents
  the very auto-compaction that would refresh them and continue the task. The skill now
  states it outright: stopping near the ceiling strands the task; keep producing work
  until Claude Code auto-compacts, then resume from the handoff. Removed "context at the
  compaction floor" from the legitimate stop-conditions list — only quota (rate-limit)
  ever justifies stopping; context never does.
• Non-redundant handoff cadence (throughput). The mid-turn ctx logic no longer nags for
  a handoff once one was saved recently (kills the redundant 6%→3% re-saving). Cadence is
  now: refresh at task boundaries, plus one velocity-timed "super close to
  auto-compaction" nudge (fires once, bypasses the throttle, suppressed if already saved).
  Updates surface "handoff already saved Nm ago" and a "≈N tool calls at this pace" estimate
  (from per-tool-call context growth) so the agent knows it's captured and how close it is.
• Context handoff-nudge held later — utilize context to the core. ctx_bands lowered so the
  "context getting low" nudge fires near the ceiling (~4–8% left) instead of at 25%/40%. The
  handoff is one cheap call that only needs to land before compaction, and the velocity-timed
  "super close" message remains the final safety net — so the agent uses nearly all its context
  before prepping a handoff, rather than slowing down early.

Added

• Current time + timezone awareness. Every prompt stamp now leads with the user's local wall
  clock (e.g. now Thu, Jun 18, 17:53 America/Los_Angeles) so the agent can reason about
  time-of-day, scheduling, and deadlines. Labelled distinctly from quota to avoid the
  clock-time/budget conflation the stamp already guards against.
• Reset-aware quota guidance. Mid-task advice no longer says "descend / slow down" when the
  5h window will RESET before the velocity engine projects exhaustion (or a reset is ≤10 min
  away) — the real risk is running dry before the reset, not a low %. Deterministic from
  resets_at + projected_exhaustion; optimism requires a positive signal, so unknown burn
  stays cautious.

v0.4.0

Added

• Continuity handoff doc — the agent's living, canonical working-doc (T2.29, ADR-18).
  A new handoff MCP tool (the fourth write surface) lets the agent maintain an evolving
  markdown document a fresh instance reads to resume fully: mission, current state,
  progress, exact next steps, key references, decisions + why, the user's own
  directives/corrections, system/process improvements discovered, and open questions.
  Stored at ~/.headroom/continuity/<session>.md, latest-wins, capped per section, session-
  guarded, pruned after 7 days. Re-injected at SessionStart(source=compact) as a pointer +
  digest (ADR-11 style — the doc lives on disk, so point, don't dump). Richer and more
  durable than the terse checkpoint. headroom handoff [--path] prints the current doc.

Changed

• Skill: context-pressure is a write-the-handoff signal, not a stop sign. Field report:
  agents were getting "tired"/cautious as compaction approached and slowing down or stopping.
  The "near the context ceiling" section is rewritten into a handoff ritual — keep a living
  handoff doc, refresh it when context runs low, then keep working at full speed; let the
  window burn to the floor because compaction is automatic and your doc + ground truth are
  re-injected right after it. The mid-turn ctx band update now says the same. (Wording
  eval-gated per ADR-9; validating in soak.)

v0.3.3

Changed

• Skill: "a clean boundary is a checkpoint, not a stop." Agents were treating natural
  pause points (tests green, a commit landed) as stopping points even with healthy budget
  and work remaining. The skill now says to continue through, and to stop ONLY when:
  nothing valuable is left, a genuine blocker needs the user, or budget is truly dry
  (≲2%). The worth-it test brakes both ways — no premature stop, no runaway. Completes the
  anti-timidity / descent-profile arc. (Wording eval-gated per ADR-9; validating in soak.)

Docs

• Product-boundary docs: the self-evolving harness is Keyoku's, not headroom's
  (EVOLVING-HARNESS.md); headroom is the awareness/sensor layer and friction feed. The
  suggest miner was prototyped and deferred (design kept in SUGGEST.md) — thinness
  applied to ourselves.

v0.3.2

Added

• Weekly cruise control: the 7d window is paced, not just measured — pace ratio
  (used-fraction vs elapsed-fraction), sustainable %/day allowance, and projected weekly
  exhaustion, surfaced in stamps ("weekly pace is HOT… ≈5%/day sustains"), HUD
  (week 26% left ⚠hot pace), fit_check advice, and burn.weekly in the state schema.
  Agents throttle bulk work when hot and cruise the week to its reset without going dark.