v1.36.2

Fixed

• Hook timing telemetry is now bounded, private, and rotation-safe. The
  unbounded timing ledger is replaced by a locked 16 MiB active plus 16 MiB
  retained ring with 0600 files, crash-aware stale-lock recovery, bounded
  records and batches, and a disk-backed migration path for large legacy logs.
  Hook flushes run in a worker so lock waits and migration cannot freeze the
  hook event loop; contended batches are retried.
• Timing reports no longer load the entire ledger or overstate incomplete
  evidence. Readers scan retained then active data in fixed-size chunks,
  filter finite windows while reading, detect writer transactions and source
  damage, and explicitly mark dropped or unknown history as partial. Dashboard
  and status conversion rates are withheld when coverage is incomplete.
• Session-end health scans are truly bounded. Tail reads now cap bytes and
  rows, include retained timing history, preserve complete newline-free legacy
  rows, and reject malformed JSON values without suppressing independent
  health nudges.

Internal

• Periodic collection during one-time legacy validation bounds the transient
  JSON parser heap on Linux as well as macOS when migrating hundreds of
  thousands of small timing rows.
• Root tests now run in four isolated Bun workers, eliminating cross-file
  process-global registry races while cutting the local suite from roughly
  72 seconds to about 22 seconds.

v1.36.0

Added

• Community Ledger + public savings leaderboard. New /community page
  (plugin.ashlr.ai/community) shows the running total of tokens and dollars
  ashlr has saved across all developers, a cumulative-savings graph, and an
  opt-in per-developer leaderboard. Backed by two new read-only public-aggregate
  endpoints — GET /public/stats/time-series and GET /public/leaderboard —
  that expose aggregate counts and (for opted-in users) GitHub handle + savings
  only; never email, code, or prompts.
• /ashlr-leaderboard command (on/off/status) — opt in or out of the
  public leaderboard. OFF by default. When on, only your GitHub handle and
  lifetime tokens/$ saved are published (the same number /ashlr-savings shows).
  Requires a stats-syncing account (Pro/Team). The cloud stats sync now sends
  leaderboard_opt_in only when you've explicitly set it. (Brings the slash
  command count to 34.)

v1.35.4

Fixed

• Lifecycle hooks emitted an invalid output shape and errored at the end of
  every session. The Stop, SubagentStop, and PreCompact hooks each wrote
  {"hookSpecificOutput":{"hookEventName":…,"additionalContext":…}}, but those
  three events are decision-control events in Claude Code — they do not accept
  hookSpecificOutput/additionalContext. Claude Code rejected the Stop
  output on every turn with Hook JSON output validation failed — (root): Invalid input. The Stop and SubagentStop hooks now keep their side
  effects (idempotent session-log finalization; subagent rollup + background
  genome consolidation) and exit silently. The PreCompact hook is removed:
  its sole purpose was context injection, which the event cannot do, and the
  same post-compaction re-orientation is already delivered by the SessionStart
  (source: compact) hook, which can inject context. Net: two lifecycle hooks
  (was three), and no more end-of-session validation error.

v1.35.3

Fixed

• Abstract classes were not chunked at all (abstract_class_declaration was missing
  from the chunker's declaration map), so abstract class declarations were absent from
  AST-read skeletons and genome RAG chunking. Now mapped to the class kind — abstract
  classes get the same member-signature treatment (and a regression test covers the harder
  shapes: abstract methods, getters/setters, generic/async methods, static + readonly
  fields, private # methods).

v1.35.2

Added

• AST-skeleton class member signatures: ashlr__read AST skeletons now render a
  class's shape — each method's signature (with the body elided as { … }) plus
  field/property/abstract/index signatures — instead of eliding the whole class body.
  Gives the model a class's callable surface for orientation at a fraction of the tokens.
  (chunker attaches members to class chunks — additive; genome RAG unaffected.)

v1.35.1

Added

• AST-skeleton enum coverage: ashlr__read AST skeletons now include enum /
  export enum declarations, rendered in full (members are an enum's signal) rather than
  dropped. Other declarations remain signature-only with bodies elided. (chunker emits
  enum chunks — also enriches genome RAG chunking.)

Internal

• CI stability: widened the brittle stub-mode bench-orchestrate perf bounds
  (parallel-speedup floor and per-node-ms ceiling) that flaked under concurrent CI load —
  they measured process-scheduling noise, not a perf SLA. Real regressions still trip them.

v1.35.0

Three WozCode-inspired token-mechanic features — the places a best-in-class
efficiency layer earns its biggest per-call wins. All additive, fallback-safe.

Added

• AST-truncated reads (ashlr__read mode: auto|ast|snip, default auto): for large
  wired code files (.ts/.tsx/.js/.mjs/.cjs/.jsx) ashlr__read returns a SKELETON —
  imports/exports + type/function/class signatures + docstrings, with function bodies
  elided — for the file's shape at ~60–80% fewer tokens. Falls back to snipCompact on any
  parse failure/timeout; bypassSummary:true still returns the full file. New
  servers/_ast-skeleton.ts (reuses the tree-sitter chunker).
• Fuzzy edit matching (ashlr__edit / ashlr__multi_edit): when an exact search
  match fails, fall back to (1) whitespace-normalized unique match, then (2) bounded block
  Levenshtein — applied ONLY on a single high-confidence match (≥0.90 + unique margin),
  never ambiguously. Eliminates failed-edit retry round-trips from indentation/whitespace
  drift. Exact-match-first and all uniqueness/error semantics preserved. Disable with
  ASHLR_EDIT_FUZZY=off. New servers/_edit-match.ts.
• Post-edit validation loop: after an edit, the resulting content is syntax-checked
  (tree-sitter for JS/TS, JSON.parse for .json) and compared against the pre-edit state
  so only edits that introduce a parse error are flagged. Default warns inline (the model
  sees ⚠ syntax … the same turn); ASHLR_EDIT_VALIDATE=block refuses the write before it
  persists. Parser failures/timeouts never break an edit. New servers/_edit-validate.ts.

Internal

• AST/JSON parsing in the edit/read paths is fully guarded (timeout + try/catch → safe
  fallback); tree-sitter is dynamically imported in the edit path to keep subprocess
  startup fast.

v1.34.1

Hardening pass. Security audit of the backend + capability-surface hardening +
honesty/labeling on experimental features + docs for the v1.34 surface.

Security

• Backend audit (server/): no critical/high app-logic issues — Stripe webhook
  signatures are verified before any mutation, genome tenant isolation is enforced
  (WHERE id=? AND org_id=?), SQL is parameterized, routes validate with Zod, secrets
  are env-injected + AES-256-GCM enveloped. Fixes applied:
  • M1 /auth/status rate limiter was disabled by a transposed-args bug
    (windowMs/maxRequests swapped) — fixed to 20 req / 60s.
  • M2 GitHub OAuth could merge into an existing account by an unverified email —
    now requires primary && verified via /user/emails before any merge.
  • M3 comp-tier grants never expired (nothing read comp_expires_at) — now
    enforced + persisted at the getUserByToken chokepoint.
  • M4 unbounded in-memory rate-limit + LLM-cache maps — added stale-entry sweeps.
  • L3 /metrics Basic Auth now uses timingSafeEqual.
  • H1 bumped hono → 4.12.23 (bodyLimit-bypass / Vary-cache / JWT CVEs). Remaining
    advisories are transitive (ws via openai, axios via @sendgrid/mail, fast-uri via
    react-email) and need major upstream bumps — tracked, not request-path-exploitable.
• ashlr__pipe capability surface: ctx now defaults to read-only tools
  (grep/read/ls/glob); shell (ctx.bash) requires a second explicit flag
  ASHLR_PIPE_ALLOW_BASH=1. Tool marked [EXPERIMENTAL]. (Already off by default via
  ASHLR_PIPE_ENABLE.)
• Added a canonical safeParse<T> hook helper + 127 hardening tests (all existing hook
  JSON.parse sites were already guarded).

Changed

• Honest labeling of experimental/stub features: /ashlr-orchestrate clearly marked an
  experimental stub-by-default MVP; AST semantic features documented as TS-only (non-TS
  falls back to line-based); stale trialEndsAt / telemetry-route comments corrected.

Documentation

• New docs for the v1.34 surface: the four new skills + ashlr-efficient output style,
  the PreCompact/SubagentStop/Stop hooks, measured-savings mode ((API-measured) vs
  (est.)), and the benchmark methodology + flags surfaced near the headline number.

v1.34.0

"Elite" — feature parity + a code-execution token lever + provable savings. First
capability-bearing release since v1.29; v1.30–v1.33 were stabilization. Three tracks
plus release-hygiene automation.

Added

• Code-execution mode — ashlr__pipe (flag-gated ASHLR_PIPE_ENABLE=1): run a short
  sandboxed expression that calls other ashlr tools (ctx.{grep,read,bash,ls,glob}),
  filters/transforms results, and returns ONLY the distilled value — intermediate tool
  output never enters context. AsyncFunction isolation (no closure/module scope),
  deny-list, ≤2000-char exprs, 10s/30s timeout, 64KB intermediate + max_output_bytes
  caps, and single-aggregate savings accounting via a _noAccounting suppress flag.
• Three new hook events: PreCompact (inject a ≤600-byte genome-ToC + session-state
  "survival kit" right before auto-compaction), SubagentStop (roll a finished subagent's
  savings into the session log + fire genome consolidation), Stop (deterministic,
  idempotent session-accounting finalization complementing SessionEnd).
• Four model-invoked Skills + SessionStart toggles: ashlr-search, ashlr-lean-tools,
  ashlr-genome-author, ashlr-cost-refactor.
• Output style ashlr-efficient — structural density (inline code, tables, inverted
  pyramid); composable with ashlr-brief.
• Measured-savings mode — servers/_token-measure.ts derives true cl100k token counts
  from the Anthropic API (max_tokens:1, fire-and-forget, SHA-256 LRU cache). Savings now
  label (API-measured) vs (est.) in /ashlr-savings, the dashboard, status line, and
  the SVG badge once ≥10 measured calls accrue.
• Reproducible benchmark — run-benchmark.ts gains bootstrap confidence intervals,
  --validate-tokenizer (heuristic-error vs API), and a --compare A/B table.
• Honest genome retrieval — grep header shows embedMode (cold/warm/hot/…), a
  one-time cold-corpus nudge, TF-IDF re-ranking + camel/snake trigram expansion in
  _genome-search.ts (no-regression on a cold corpus), and an Ollama auto-detect hint in
  /ashlr-doctor.
• Release-hygiene automation — scripts/bump-version.ts (single source of truth across
  package.json + plugin.json + marketplace.json) and scripts/check-version-sync.ts (CI
  gate). Wired into ci.yml, release.yml, and publish.sh.

Fixed

• Version drift — manifests were stuck at 1.29.0 while tags reached 1.33.0; now synced
  to 1.34.0 and enforced by CI so it can't recur.
• Stale plugin-metadata test — pinned to "31 slash commands" since PR #77; aligned to
  the canonical 33.

Internal

• _stats.ts schema extended additively (optional tokensSavedMeasured/measuredCalls,
  no schema-version bump, back-compatible).

v1.33 — Audit-driven polish + multi-host portability

18 PRs since v1.32 (#88-#105). A strategic audit pass identified 3 fatal first-touch UX cliffs + 1 underexploited competitive opportunity (multi-host MCP post-RooCode shutdown). All 4 shipped + Q1'27 wk 7-12 orchestration work landed.

UX cliffs closed

• #102 first-call savings projection — see your extrapolated annual savings on your very first ashlr call. Closes the audit's "$0.001 saved → user closes terminal" cliff.
• #104 Bun ≥ 1.3 prereq — bold callout above the install line. Closes the silent-install-failure cliff.
• #105 wizard restart enforcement — loud RESTART REQUIRED block + SessionStart hook detects missed restarts and warns. Closes the "wizard finished but plugin appears broken" cliff.

Multi-host portability

• #103 multi-host MCP — ASHLR_MCP_HOST env var gates Claude-Code-specific paths. Server now runs cleanly in Cline, Claude Desktop, OpenAI Codex CLI. Setup snippets at docs/multi-host-mcp.md. Captures the post-RooCode-shutdown OSS replacement window.

Q1'27 wk 7-12 orchestration

• #83 cross-session discovery propagation aggregator — surfaces compounding knowledge across runs
• #78 multiplayer WAD-D breakdown — per-segment + top-mover indicators
• #81 WAD-D historical drilldown — date-range picker + sparkline grid
• #82 session replay UI v0 at /admin/sessions
• #94 orchestration telemetry table + dashboard panel
• #97 orchestrator overhead bench — measured ~5-30ms/node, 2.8× parallel speedup
• #98 /ashlr-orchestrate-status — inspect past runs
• #99 central token quota accounting — orchestration_usage table + admin endpoint
• #101 retry-with-backoff + handoff context budget — 3 attempts, exponential backoff, 8KB handoff cap

Fixes

• #96 pre-existing tsc errors + permanent dashboard projection date-drift fix via clock injection
• Discovery propagation aggregator wired into the daily WAD-D cron

Strategic artifacts

Four strategic audit docs filed at ~/.claude/plans/:

• audit-code-quality.md — production-ready; 3 critical findings mitigated in practice
• audit-first-touch-ux.md — 5 minutes not habit-forming; value emerges Day 3-5
• competitive-landscape.md — positioning + RooCode shutdown opportunity
• groundbreaking-bets.md — 5 post-northstar bets; Code Sandbox is the most ambitious

Internal

• 36 PRs merged across the v1.31 → v1.33 arc (#66-#105)
• ~720 new tests
• 2 release tags (v1.31.0 + v1.32.0 + v1.33.0)
• 4 issues filed + closed (#69 + #84 + #85 + #86)
• WAD-D + orchestration ingest both live in production