stratless 0.3.5 — the worker: the work moves off your terminal

Phase 2 of the cold-start build. The machinery no longer lives inside the command you typed — it
lives in ONE background worker that commands merely wake. Nothing runs when there is nothing to
do; there is no daemon.

Changed

• update is now a doorbell. It wakes the background worker and — in a real terminal —
  watches it, printing the same progress lines as before. The difference shows only at the edges:
  Ctrl-C (or closing the terminal) detaches the display; the work continues and the profile
  still lands. The detach message prints the whole kill ladder: watch with status, stop
  everything with stop. The after-session hook rings the same doorbell and returns instantly.
• stop now stops a RUNNING refresh too, within seconds. The off switch means spending halts
  now — not after the current build finishes. Killing never wastes what was spent: every judgment
  already made is banked, and restarting re-reads at most one chunk.
• status shows a live worker (running now judging 12/31 · pid N) and labels the last
  run honestly (stopped by you, or failed with the reason).
• Kill-safe to one chunk, for real (C3). Judged verdicts are now banked per streamed session
  as they land — a crash, sleep, or kill loses at most the twelve-turn chunk in flight, never the
  whole batch. (The Phase 2 gauntlet test caught the old behavior losing the entire batch.)
• Typos refuse instead of quietly running something else. stratless update --npw used to run
  a plain update while you believed you had forced a rebuild. Unknown flags and stray arguments
  now exit loudly, with a did-you-mean — and a mistyped command (stratless updat) gets the same
  courtesy, suggesting the nearest verb.
• The borrowed model pin is absolute. A stage pinned to sonnet (mining, writing, grading) or
  haiku (judging) now runs on that model or REFUSES — it never silently falls back to your account
  default. Left unchecked, a failed pin could land the priciest stage on the priciest model (if you
  default to Opus) at frontier rates, invisibly. Override stays yours via STRATLESS_SYNTH_MODEL;
  a CLI without JSON output still answers, but on the same pinned model.

Added

• Every run hands you its receipt. A finished update closes with what it actually spent —
  this run: 1.2M tokens · ≈ $0.21 at API rates · claude-haiku-4-5 ×31 · claude-sonnet-5 ×3 —
  tokens first, models by their ground-truth names (which model RAN, not which was asked for).
  Refused and stopped runs get their receipt too; a run that spent nothing owes none. status
  keeps the most recent run's receipt (last run spend), so the hook's silent spends stay
  readable. Announced before, metered during, accounted after.
• The artifact-shape lint (the second half of the tool-less-borrow guarantee, pulled forward
  after a chatter reply was loaded as a real HUMAN.md in the wild): a profile that opens with
  assistant chatter, or a pattern-era profile without its section headings, is REFUSED like an
  invented number — nothing malformed ever loads.
• The flat-memory walk (C1): the archive is read one transcript at a time, newest first —
  memory stays flat whether your history holds 200 exchanges or 20,000. This is the read the
  cold-start release (0.4.0) will stand on.

---

Published to npm over Trusted Publishing (OIDC, no token) with Sigstore provenance — verify with npm audit signatures.

stratless 0.3.4 — the plumbing: what the cold-start build stands on

Phase 1 of the cold-start build (spec §12). No new features — the release is eight acceptance
criteria passing, three of them bugs the Phase 0 measurement sitting caught in the shipped CLI.

Fixed

• "MOST RECENT" meant oldest. Since 0.3.1, the writer's most-recent evidence block received
  the OLDEST 25 judgments of a newest-first list — every pattern-era profile had its
  current-direction evidence backwards. Now selected by timestamp, newest first, regardless of
  caller order, with a test that pins the ordering (C10).
• The meter's blind spot is closed. A call that degraded to the plain-text fallback used to be
  recorded at zero cost, and a call that lost its model pin silently ran on your account's default
  model at frontier rates. Both are now counted — status shows fallback calls: N unmetered · M ran on your default model the moment either exists, and the ledger records which model actually
  ran (C11).
• The borrow is tool-less. Every borrowed claude call — judge, mine, audit, grade, write —
  now runs with --tools "". The borrowed model sometimes took the writer prompt literally,
  attempted a real file write, and returned permission chatter instead of your profile; it no
  longer has hands (C9).
• A failed call's error message can no longer be mistaken for an answer. A broken model pin
  exits 0 with an error envelope whose prose ("There's an issue with the selected model…") the
  judge would have cached as a verdict and the writer would have reasoned from. The envelope's
  is_error flag is now honored: the call is refused and the ladder advances to the next metered
  rung instead. (Found by this release's own review pass, verified against the live CLI.)

Changed

• Every store is written atomically (temp + rename): judgments, patterns, state, usage,
  renders, settings, HUMAN.md, CLAUDE.md. A crash mid-write leaves the old file or the new one,
  never a torn one (C2). Writes follow symlinks and preserve the file's mode — a dotfiles-repo
  link to your CLAUDE.md or settings.json survives, and your chmod stays yours.
• A damaged spend-cache refuses loudly instead of re-billing you. A corrupt judgments.json or
  patterns.json used to read as "empty" — which would silently re-spend your entire history on the
  next run. Now every spending command refuses with the file's path and how to move it aside, and
  status labels the cache unreadable instead of showing 0 (C2).
• One spender at a time. update, and a building profile/report, take a lock; the
  after-session hook and a hand-run update can no longer race each other's writes over the
  judgment cache. A second runner says another stratless run is active and exits clean; stale
  locks from crashes are detected (with PID-reuse protection) and stolen (C4).
• Rate limits bend the run, they don't break it. A streamed batch that hits a 429/overload now
  backs off exponentially and retries the remaining items in a fresh session, bounded, losing
  nothing — a rate-limit storm costs minutes, not evidence (C6).

Added

• The stopwatch. Per-turn, per-stage, per-run wall-clock recorded in state.json on every
  spend — so the cold-start door (0.4.0) can quote minutes measured on YOUR machine, never typed.
  Includes a pure ETA that refuses to estimate any stage it hasn't measured (C8).
• The detached-spawn primitive the Phase 2 worker will stand on: spawn a process that survives
  its parent, with the absolute path to your claude captured at spawn time (C5). Unused by any
  command yet — built and tested ahead of the worker.

---

Published to npm over Trusted Publishing (OIDC, no token) with Sigstore provenance — verify with npm audit signatures.

stratless 0.3.3 — looking is free, evidence is readable, and the tool talks properly through a pipe

Changed

• Looking is free. stratless profile and stratless report now print the last built
  rendering instantly, at zero spend, under a header carrying the build's own date and numbers.
  Only update — and an explicit --now, or a first-ever look — spends. One word keeps one
  meaning: --now = spend now, skip nothing.
• The bill is announced, never discovered. Before any run makes a fresh read, it says so
  first: about to read 37 new exchanges on your own claude (each read once, cached forever).
• Clean pipes. Output through a pipe, with NO_COLOR set (and non-empty, per the spec), or
  under TERM=dumb carries zero escape codes — bold and dim stripped too, stricter than the
  NO_COLOR spec on purpose. stratless status | grep loaded finally behaves.
• -h now works alongside --help and help; the help screen leads with the getting-started
  line and ends with the docs link; report closes by pointing at its evidence.

Added

• stratless receipt <n> — prove any claim. patterns now numbers every claim and each one
  says prove it: stratless receipt 3. The receipt command dereferences a claim back to the raw
  exchanges behind it — what you asked, what it said, how you reacted — re-read from your own
  transcripts (archive first). Hash prefixes work git-style. Free; zero model calls. And when a
  receipt points at a transcript the 30-day cleanup deleted before your archive existed, it says
  exactly that: the honest failure is the init lesson.
• A version check you consented to, and nothing ambient. stratless status --check asks npm
  for the latest version — user-initiated, on-screen, for everyone; plain status stays fully
  offline. --auto users (who explicitly armed background activity) get a cached once-daily check
  riding the background refresh, disclosed at arming time. Plain-init users make zero registry
  calls, ever — and the test suite proves the unarmed path can't reach the network. The privacy
  page now tells this tiered truth in full.

---

Published to npm over Trusted Publishing (OIDC, no token) with Sigstore provenance — verify with npm audit signatures.

stratless 0.3.2

The self-correction release. Until now the profile asserted its confidence; from here it earns it.

Added

• The grader. Every pattern is a dated prediction ("this will keep happening"), and each
  rebuild now grades it against the window's new evidence: confirmed (new receipts landed),
  silent (the topic never came up — never counted against a claim), or surprised (evidence
  shows the opposite; the mistake carries its own receipts). A separate mind on Sonnet, streamed;
  strict rules ("absence of the behavior is NOT the opposite of the behavior") tuned in three
  controlled A/Bs on identical evidence.
• Balanced revision. One surprise costs confidence, never standing; a second inside two weeks
  flags the statement back to the miner: revise it to fit ALL the evidence, or retire it (retired
  claims release their evidence for future patterns; revised ones re-earn their audit from zero).
  Validated on a real month-apart backtest: June-mined claims graded against July caught exactly
  what they should — statements that overreached ("always X") meeting the fuller truth.
• The report tells you when the file was wrong. New surprises appear in stratless report as
  honest updates, scaled to their weight — a first counter-example gets a clause, a claim under
  revision gets the story. HUMAN.md itself stays clean: the AI reads current truth, you read the
  diary.

Changed

• HUMAN.md speaks only in facts about you. Pipeline bookkeeping (audit tallies, trend labels,
  confidence classes) no longer appears in the profile — the writer uses it to weigh what to say
  and how firmly, and says it in calibrated language instead. Real frequencies of your own behavior
  stay when they carry meaning.

Fixed

• Audit tallies were silently resetting on every mine (pattern ledgers now survive re-aggregation).
• An async pipe error in the streaming layer could kill a whole run mid-batch (found by the
  backtest; completed turns now always survive and the remainder falls back).

---

Published to npm over Trusted Publishing (OIDC, no token) with Sigstore provenance — verify with npm audit signatures.

stratless 0.3.1

The streaming release. 0.3.0's dogfood measured where the money actually goes: every judge and
audit call re-booted the whole Claude Code harness — ~30k tokens and ~10 seconds of startup for a
one-line question, 99% luggage. 0.3.1 pays the toll once.

Changed

• Judging and auditing now stream through ONE claude process per batch. The instructions ride
  the session's system prompt (sent once, cached); each judgment or audit is a turn. Measured in
  prototype: ~3.2x cheaper per verdict and no per-call boot; sessions rotate every 25 turns to
  bound context growth. The old per-call path remains as the automatic fallback (older CLIs,
  mid-batch failures) — the ladder is stream → per-call → silence, and a failed turn is skipped,
  never guessed. Streamed prompts carry a <stratless-…> sentinel so the pipeline can never parse
  its own sessions as your conversation.
• The profile is SECTIONED — the six kinds as fixed headings (WHAT THEY KNOW · HOW THEY THINK ·
  HOW THEY WORK · DIRECTION · FAILURE SIGNALS · TRIGGERS), only sections with evidence appear,
  under 350 words, with a humanmd/v1 schema marker in the file. HUMAN.md starts looking like the
  protocol it is becoming.
• Time is narrated. The trend and window every pattern carries now reach the prose: "has held
  for two months, lean on it", "faded through July: who they were". And the miner now sees each
  judgment's local time tag (computed by code, never raw timestamps), so time-conditioned patterns
  ("terser in the morning") can finally emerge.
• report joins the pattern era — the human's mirror is now reasoned from the audited
  patterns and centers the trajectory: what stopped, what holds, what's new. It also gains the
  numbers-lint: both renderings now refuse to deliver an invented number.

---

Published to npm over Trusted Publishing (OIDC, no token) with Sigstore provenance — verify with npm audit signatures.

stratless 0.3.0

The learning release. Until now the profile was re-derived from raw judgment lines on every build
and the derivation thrown away; nothing accumulated. 0.3.0 makes the derivation a persistent,
auditable artifact — the profile stops being an impression and becomes a derivation from evidence.

Added

• The miner. A new pass turns your judgments into ~/.stratless/patterns.json: named
  regularities, each carrying a real count, a time window, a trend (rising/steady/fading), a
  stability class, and — the part that matters — receipts: the exchange hashes that witnessed
  it. Every claim traces to replies that could have gone the other way; no receipt, no claim. The
  split is strict: the model names, the code counts — counts, dates and trends are computed
  deterministically and the model is never shown a number it could hallucinate. Categories emerge
  from your data (a pattern needs 5+ receipts to exist; anecdotes stay candidates); they sort into
  six fixed kinds — what you know · how you think · how you work · direction · failure signals ·
  triggers — and anything that fits none is kept unsorted, never forced.
• stratless patterns — every claim with its count, trend, audit tally, and receipt trail.
  Your profile's evidence, inspectable. Free; add --all for candidates below the evidence bar.
• The auditor. A separate mind — never the one that wrote a statement — re-checks every newly
  assigned receipt against its pattern ("does this judgment actually SATISFY the statement?").
  Evicted evidence is removed by code and admission is re-decided by arithmetic. Audited once,
  ever, per receipt.
• The numbers-lint. Every numeral in a built profile must already exist in the evidence the
  writer was shown. An invented or rounded frequency is a refused build — the old profile stays
  loaded, nothing lies. A wrong frequency is a lie wearing precision; now it's a build failure.
• Dynamic aperture. The judge's view of each exchange is sized from your own history (p90 of
  your real lengths × 1.2, clamped), reading head + tail of long turns — the plan and the
  conclusion — instead of a fixed head-cut sized on the author's machine. Recorded in
  state.json, visible.

Changed

• Judgments are structured (v2). Verdict / topic / behavior fields, JSON output validated in
  code (malformed = refused, never guessed). The pipeline version now lives in the cache key, so
  this release re-judges your window gradually, under the normal per-run budget — an upgrade
  never re-spends your backlog at once.
• The profile is written FROM the patterns — audited claims with real numbers, plus a small
  recent sample for freshness. The writer's input stays the same size whether you have 200
  judgments or 20,000: the learning architecture is the cost-control architecture.
• The parse cap rose 4,000 → 8,000 chars per field (inside the v2 re-judge, so the identity
  churn was free). Long assistant turns keep more of their real tail.

---

Published to npm over Trusted Publishing (OIDC, no token) with Sigstore provenance — verify with npm audit signatures.

stratless 0.2.4 — the cost release: synthesis when due + an honest meter

Changed

• update now rebuilds the profile only when it's due — the cost release. The synthesis is the
  expensive read (~32 judge calls' worth, measured on real usage); it used to run after every
  session, even for a handful of new exchanges. Sessions now accumulate: the profile rebuilds after
  25 new judgments (STRATLESS_SYNTH_EVERY overrides), when it's over a week stale and anything
  new arrived, when no profile exists yet, or on stratless update --now. A skipped rebuild
  still judges what's new and still guarantees the profile stays loaded — the skip is invisible,
  only the cost is missing.
• The profile is written by Sonnet, pinned — same deliberate read, ~4–6x cheaper. Synthesis used
  to ride whatever your default model is; a frontier default spent ~$0.62 API-equivalent per rebuild
  (thinking billed at frontier rates), which was ~80% of everything stratless spent. Sonnet thinks
  just as hard about who you are for ~$0.10–0.15. STRATLESS_SYNTH_MODEL brings your default back
  if you want it.
• The judge now reads the END of a long assistant turn — the part you actually reacted to. Long
  turns were cut from the head, so the judge read the opening preamble and then your reaction to a
  conclusion it never saw. Measured on real history: 83% of turns were longer than the judge's view.
  Already-judged exchanges are not re-read; only the truncated ones re-judge, gradually, in the
  background budget.
• Reactions that carry no signal (none — logistics, a thank-you) no longer reach the writer.
  They stay cached like everything else; they just stop diluting the profile.
• profile now says where the load stands. The split is deliberate — profile and report
  look, update loads — but profile used to end in silence, leaving the impression the
  printed profile was live. It now closes with the honest state: not loaded yet — load it into your assistant: stratless update, or the loaded path and how to refresh it.
• Every printed next-step now works in your shell. Run stratless via npx and the hints say
  npx stratless profile; run it installed and they stay bare. (Following npx stratless init
  with the suggested stratless profile used to hit command not found.)
• stop points back the right way. Its parting hint now names what actually reverses it
  (stratless update to reload, stratless init --auto for background refresh) — plain init
  no longer arms the refresh since 0.2.3, and status said the same stale thing.

Fixed

• The usage meter no longer under-reports by ~2,000x. It recorded only plain input tokens and
  dropped the cache tokens where the real consumption lives (~17–24k tokens of Claude Code harness
  overhead per borrowed call) — the ledger showed 484 input tokens where reality was over a million.
  status now reports real tokens first (a subscription spends quota, not dollars), the
  API-equivalent cost clearly labelled, and the judging / profile-build split. Existing tallies
  survive the upgrade untouched.
• A hand-edited ~/.claude/settings.json no longer crashes init. Malformed JSON (the classic
  trailing comma) used to become a raw stack trace — and could not be distinguished from a file worth
  overwriting. init now refuses with a clear message and touches nothing; stop treats it as
  "nothing to remove".
• init --auto warns when the background refresh can't actually run — the hook calls the bare
  stratless, which an npx-only install doesn't put on your PATH. It now says so and names the fix
  instead of failing silently every session.
• The "report this" links printed when stratless refuses on format drift pointed at the repo's old
  name and led to a 404. They point at the real repository now.

---

Published to npm over Trusted Publishing (OIDC, no token) with Sigstore provenance — verify with npm audit signatures.

stratless 0.2.3 — no more history-load hangs + opt-in auto-refresh

Fixed

• profile / report / update no longer load your whole history into memory. They read only the recent window (newest transcripts first, then stop), so a large archive can no longer exhaust memory and hang the machine. Previously the load read and parsed every transcript before keeping the last 200 — the 0.2.1 "bounded cold start" bounded only the judging, not the load. Measured on an 833 MB / 2,292-file archive: ~165 MB and under half a second, versus the old out-of-memory hang.

Changed

• The after-session auto-refresh is now opt-in. Plain stratless init sets up the archive and the reaper but installs no background hook; run stratless init --auto to have your profile rebuild itself after each session. stratless stop still turns it off. A tool that reads your history should not silently arm a background job on every session.

---

Published to npm over Trusted Publishing (OIDC, no token) with Sigstore provenance — verify with npm audit signatures.

stratless 0.2.2 — status command + cost tracking

Added

• stratless status — stratless's own state at a glance: whether the after-session refresh is on, whether your profile is loaded, how many exchanges have been judged, when it last refreshed, and the running total it has spent on your borrowed claude. It reads locally and spends nothing. (stats still counts your assistant's activity in a project; status is about stratless itself.)

Changed

• The generated profile is now em-dash-free at the source — the synthesis prompt avoids em and en dashes, and a deterministic pass strips any that slip through, so HUMAN.md stops carrying the one punctuation mark that reads as machine-written.

Fixed

• Housekeeping: bumped pnpm/action-setup off its deprecated Node 20 runner in CI, and removed a stale workflow comment and a couple of stray files.

---

Published to npm over Trusted Publishing (OIDC, no token) with Sigstore provenance — verify with npm audit signatures.

stratless 0.2.1 — bounded cold start + --version

Changed

• Bounded cold start. The profile is now built from a recent window (the newest ~200 exchanges) rather than the whole backlog — it converges there anyway. Every run and after-session refresh stays fast and safe at any history size. (Previously, a full pass over a large backlog could stall or crash the machine.)

Added

• stratless --version / -v — print the installed version.
• CHANGELOG.md (Keep a Changelog), shipped in the package.

Removed

• --backfill — the unbounded "judge everything now" flag that could take a laptop down. The bounded window replaces it; it drains a little each run and through the after-session hook.

---

Runs on your machine. Reads your own history. Nothing leaves. Published with npm provenance — verify with npm audit signatures.