waybill 2.1.0

The launch-readiness release: the meter now sees everything Claude Code
writes, the security prose says exactly what the mechanisms do, and the
launch numbers come from the engine metering the session that built this
release. Minor per the migration policy: every
schema change is an additive optional field; no existing event re-derives
differently; no migration.

Historical totals will grow on upgrade — that is the fix working.
Subagent transcripts were previously invisible to the meter (on one
measured real session, 28.6M of 41.8M tokens went uncounted). The next
mine/meter --all discovers them as new sessions with their own
checkpoints — no blanket re-meter, no METER_LOGIC_VERSION bump, and
your existing sessions' events are untouched.

Fixed

• Subagent transcripts metered (E-02) — the walk, the SessionEnd
  queue path, and --transcript now discover
  projects/<proj>/<session>/subagents/*.jsonl. Each meters as a
  session of its own under a composite id (<parent>:agent-<id>), with
  its own receipt, checkpoint, and conservation check; a whole-session
  pin on the parent covers its subagents; user-facing session counts
  group by root session; packs include a session whole together with
  its subagent events.
• Three overclaiming sentences rewritten to the mechanisms (E-01) —
  the pack README no longer claims a backdated estimate "fails the
  seal" (the seal proves no edits since a copy was shared, not when the
  estimate was written); the README states conservation is an
  internal-consistency guarantee that cannot detect undercounting at
  the source; the architecture doc describes the escrow payload order
  as a field-shift/delimiter-injection defense, not "structurally
  impossible". Write-time ordering is now enforced where a wall clock
  exists: append refuses a future-dated logged_at (5-minute skew),
  stamps ledger entries with an additive appended_at witness, and
  verify discloses — as a warning, never a failure — a pre-registered
  estimate written more than 48h after its claimed logged_at.
  Backfilled sync facts entries carry no estimate and no stamp, so they
  cannot trip it. External time anchoring is roadmap, recorded.
• The demo renders everywhere (E-03) — base opacity 1, so renderers
  that don't run CSS animation in <img> (Firefox on GitHub, previews)
  show the complete static transcript; animated playback cold-opens on
  the payoff frame and holds the finished transcript instead of fading
  to nothing.
• Meter gaps reach verify and every pack (E-11) — sessions whose
  transcripts were pruned or unreadable before metering are disclosed
  as verify warnings and travel in packs (gap_sessions in pack.json,
  a "what the totals do not contain" line in the pack README) instead
  of being silently absent behind green.
• SessionStart runs no engine work (E-12) — the miner precomputes
  the pacing/renewal/first-run notice at mine time (which is when
  thresholds actually cross) into rollups/next-notice; the hook is
  pure shell: serve the file, consume it, exit. The not-initialized
  nudge is owned by the hook itself. pace --notice is unchanged.
• First receipt contains tokens (E-04) — init meters every
  existing transcript (subagents included) with a progress line, so
  months of history are visible immediately; the README shows the real
  receipt this produced for this very repo (E-05), and the install
  block says to restart Claude Code (E-07).
• PAT guidance unified on fine-grained read-only (E-13); the
  waybill-not-on-PATH answer documented (E-08); good first issue
  labeling fixed (E-09); bug intake asks nothing users can't do (E-16).

Added

• query cache + the cache skill (S-04) — what the bill is
  actually made of: volume by class with the 5m/1h write split, the
  cache-read share, per-model rows, and a derived effective-vs-list
  cost view (labeled list_price_equivalent_derived, net of write
  premiums, coverage disclosed). Triggers on "why is my bill like
  this"; trigger + negative evals included.
• Multi-repo session disclosure (E-14, warn-only) — receipts record
  every working directory seen (additive cwds, present only when >1);
  verify warns that such a session's spend is attributed via the
  first cwd only. The per-turn split is future work, tracked.
• A warnings channel in verify — findings can carry
  severity: "warning": disclosed by verify and status, never
  fatal, never blocking packs. Used by the write-time-lag, meter-gap,
  and multi-repo disclosures.
• Evidence-tier mix in status (S-05) — one line: points shipped,
  points with pre-registered estimates, points facts-only.
• Capability phrasing (E-15) — "what can waybill do", "waybill
  help", "show me my dashboard" trigger the right skills, with evals.

Changed

• Positioning reframed to the defensible claim (S-01) — receipts,
  not estimates; the absolute "unserved / no tooling" language is gone
  repo-wide. The dated, evidence-tiered field survey lives in
  docs/positioning.md; the README first screen
  leads with the install unit, the real receipt, and the cache-read
  share (S-02). The launch post carries only engine-generated numbers,
  unattributed share included (S-03).
• Walks share one stream read (E-10) — mine, meter --all, and
  init's first-run pass load the four streams once per walk instead of
  once per transcript (byte-identical output, tested).
• ROADMAP, DECISIONS, and VALIDATION moved under docs/; the
  pre-2.0 changelog folded into a details block; .mailmap normalizes
  tool-default commit authorship (E-06).

"Delivered": the free side complete — the final release of the committed
path (ROADMAP). The major version marks scope completion,
not breakage: the event schema is unchanged (v2, frozen since 1.0), every
addition is additive, and 1.x ledgers need no migration. From here,
changes are upkeep, not construction.

Added

• Azure DevOps tracker adapter — REST/WIQL work items → WorkItems:
  numeric work item ids as keys (AB#123-style branch refs match),
  StoryPoints or Scrum Effort → points (never invented), iteration
  path tail → sprint, ClosedDate/terminal states → done, work item
  type → work_type. Own-data defense-in-depth reads
  System.AssignedTo.uniqueName (an email) against the identity map's
  existing git_emails — no new identity field needed.
  Conformance- and own-data-tested on a realistic fixture
  (--tracker azure-devops). Honest limit: live end-to-end run not yet
  performed; the adapters table says so.
• Bitbucket Cloud git-host adapter — merged pull requests →
  MergedChanges: repo from the destination repository, branch/title keys,
  and — documented approximation, verbatim leaf — updated_on of a
  MERGED PR standing in for the merge time (the 2.0 PR object exposes no
  merged_on). No closing-keyword linkage on Bitbucket, so closes is
  omitted, never faked. Own-data via new optional
  identity.json.bitbucket_username. Conformance- and own-data-tested
  (--git bitbucket).
• The Windows story (docs/windows.md) — the
  engine is pure Node and runs everywhere; hooks need only a bash on
  PATH (Git for Windows, WSL); and because metering is retroactive and
  idempotent, a machine where hooks never fire loses nothing — scheduled
  mine --all is a full substitute. Retention is the one real caveat,
  and status already checks it. Honest limit stated: reasoned port,
  native end-to-end report welcome.
• The OTel live-export recipe (docs/otel.md) —
  telemetry env vars → local OTLP collector with a file exporter →
  waybill meter --otel backfills pruned-transcript sessions. Fallback
  semantics restated: fills transcript-less sessions only; the
  transcript stays the source of truth.
• The self-verifying release gate (scripts/release-gate.sh,
  npm run gate, new CI job) — the release's claims about itself are
  checked mechanically: VALIDATION's headline test count must match the
  suite that actually runs, the released version must have CHANGELOG and
  ROADMAP coverage, every relative doc link must resolve, and every
  skill must have a trigger eval.
• status --fast — skips the full-ledger verify re-hash (the one
  expensive step on a large home); the verify line reports "skipped",
  never "ok".

Changed

• Architecture review closed out — all eight standing
  recommendations from the 1.5.0 review now carry a recorded disposition
  in docs/architecture.md: done in code (time
  module v1.6; own-data kit v1.7; lock-takeover post-rename verification,
  status fast mode, release gate, standup↔spend agreement golden — all
  v2.0), or deliberately closed with rationale (the shared strict flag
  parser src/cli/flags.ts is adopted by the simple read-side commands,
  while tested bespoke parsers on the write paths stay — strictness is
  enforced by regression tests either way). The pricing-honesty
  vocabulary (unpriced / repriceable / unidentified) now has one
  owning definition there.
• Lock takeover hardened — after the atomic rename that reaps a
  stale miner lock, the reaper re-reads what it actually claimed; a lock
  that turned live inside the race window is restored via
  exclusive-create and the takeover yields.

Earlier releases — 1.8.0 back to 0.1.0 (the 0.x build-out and the 1.x committed path, all 2026-08)

CHANGELOG · VALIDATION · DECISIONS