v1.70.0

What's Changed

Added

• Subagent control verbs. subagent_control stops, cancels, and re-dispatches live children by
  name; cancel finalizes through the standard delivery funnel so exactly one result lands and the
  scope lock releases.
• Stalled-child recovery and atomic panes. Stalled children are detected and recovered; panes
  launch atomically, dead multiplexers heal, completion signals travel over a wait-for channel with
  an exit sidecar.
• Queueing past the cap. Spawns beyond the concurrency cap queue without blocking the turn.
• Declared file-scope locks. A spawn's declared files: scope queues conflicting spawns FIFO
  behind the owner; read-only and unscoped spawns never queue; locks release at result delivery and
  survive pane-loss restart.
• Result records and prior-work injection. Results carry files_changed from a spawn-time git
  baseline; append-only subagent history (rotated to the last 500 lines) survives result-inbox
  drain; a scope overlapping recent history injects capped prior-work context into the task.
• Live metrics and baselines. Per-child live metrics with p90 history baselines and
  duplicate-spawn warnings; dashboard subagent health cards, deep links, and a drill-in session
  modal; a wave-metrics extractor script; the dashboard UI is built in bootstrap and after-pull.
• Fail-closed gate for children. Destructive confirms fail closed inside subagents (no dialog
  can be answered there), and the permission gate is re-enabled in every sandboxed child.
• Roster at 14 specialists. Six new profiles — architecture review, security review, frontend,
  Rust, product briefs, session audit; fixer absorbed into worker; every profile hardened against
  logged failure modes.
• Issue-orchestrator facts-file handoff. One facts file per issue group, written once; workers
  re-read only the files it names; routing splits Rust, browser-UI, and general work onto matching
  agents.

Changed

• Doctrine lives in docs/subagents.md. Orchestration rules moved out of AGENTS.md; the
  roster table plus agents/ frontmatter are the sync surface; AGENTS.md keeps only the opt-in
  gate and a pointer, shrinking every session's always-on context.
• Faster delivery paths. Single-parse result delivery, deduped session reads with cached agent
  discovery, byte-offset budget reads.
• Docs ground the workflow. README subagent-orchestration section; scope locks, control verbs,
  and executor doctrine documented; token-aware delegation gate enforced in the agents docs;
  prompts forbid touching the default branch.
• Default thinking level lowered to high; repo-local .pi/ runtime state ignored.

Fixed

• Subagent state isolated per project. History and inbox no longer leak across projects; stale
  inbox records reach unrelated sessions as context-only.
• Delivery hardening. Watcher aborts mid-poll settled; multiplexer stderr kept out of failure
  wraps; history crash residue tolerated on append; suite type errors repaired; test suites
  isolated from the real durable inbox.
• piread null handling. Call labels screen serialized null hits; explicit null accepted for
  optional params.
• Issue orchestrator correctness. Refined base sanity check and issue processing; merge order is
  canonical everywhere — PR creation order, oldest first.

v1.69.0

What's Changed

Added

• Fused locate+read (piread --hits). One call returns merged context windows around every
  matching line plus pointers to hits past the budget: 72,445 B / 27 calls against the built-in
  read's 399,683 B / 11 on the 27-task lookup corpus (0.18x the bytes).
• Wider default read budget. A read now carries 100 lines / 8 KB before a gap marker (was
  40 lines / 2 KB), so ordinary windows arrive whole and follow-up reads are rarer.
• Read budget and locate bounds are generated, so the Rust tools, the gate and the docs read
  the same numbers.

Changed

• Every native tool is Rust-only. pibash, piread, piedit and pidx register unconditionally and
  never fall back to a built-in executor; a missing or broken binary is a tool error plus one
  stderr line instead of a silent downgrade.
• Built-in tools are a configuration choice. PI_NATIVE_READ, PI_NATIVE_BASH,
  PI_NATIVE_EDIT and PI_NATIVE_PIDX are gone; use pi's read/bash/edit by leaving the
  extension out or by setting the gate's nativeReadFirst / nativeBashFirst rules to off.
• Read coverage comes from structured windows. The gate certifies a fused read from
  details.locate (printed windows, complement recorded as unread gaps) and seeds locate-first
  anchors from the reported hit lines; the read ledger stores the same complement.
• Images read as text. piread serves image and other binary files as lossy text like any other
  content, and the built-in image route is gone; the tool no longer points at a read tool that
  may not exist.

Fixed

• npm run dashboard boots again. The server had lost its configuration block, so it died on
  ReferenceError before listening; it now reads its sessions tree, pidx cache, gate config, poll
  and active windows, honours PI_DASH_PORT / PI_DASH_SESSIONS_DIR / PI_DASH_SESSION_BUDGET,
  and logs the port it actually bound.
• doc:check reports all four families clean. The checker now scans dashboard/scripts/, waives
  git subcommand flags, and every documented environment key resolves.
• Fused hit pointers and structured hits agree. details.locate.hits carries the printed hits
  plus the footer pointers, so locate-first anchoring no longer misses lines the footer named.

v1.67.0

What's Changed

Added

• Subagent budgets. Every subagent now runs under finite ceilings — wall-clock seconds, assistant
  turns, cumulative USD cost — with a clean stop and a budget_exceeded result when breached.
  Ceilings resolve from agent budget: frontmatter, budgetDefaults config, or builtin defaults;
  spawn-time overrides may only lower them.
• Budget persistence on resume. A resumed subagent inherits its spawn-time budget: the wall
  clock restarts while turns and cost re-meter from the preserved session file.

v1.66.0

What's Changed

Added

• Self-learn lessons store and skill routing. Verified lessons route to the home that owns
  them: an existing skill whose domain covers the topic is updated through skill-manager, while
  ownerless topics land in skills/self-learn/references/lessons/ — one file per topic in a
  trigger, mistake-shape, fix-code-block, evidence format that rejects prose-only entries.
• Deployed code shapes in map-integration. references/realtime-plotting.md gains the
  live-trace dashboard shapes behind its rules: the session-configured MQTT connection hook
  (latest-ref callbacks, topic-set-keyed reconnects, bounded credential re-auth), the interval
  batch buffer with its flush gate, the per-entity batch_update store merge, the deck.gl
  overlay via useControl, and style-gated throttled viewport reads.

Changed

• Anchor Guard names every uncertified edit span. A multi-edit batch that misses
  certification reports each failing edits[i] with its target line range and the exact
  piread call that certifies it, replacing one opaque failure; the case corpus pins the
  message format.
• Read/edit prerequisites restated. SYSTEM.md scopes anchors to the current turn, requires
  a search to name first-time paths (script-created files included) before piread, demands a
  region re-read after half-failed batches or earlier-turn changes, pastes printed lines
  verbatim into retries, resumes stopped reads at the named offset, and adds a context-budget
  rule (~150K tokens) that prefers compaction over continuing. AGENTS.md drops the stale
  commitlint filename, marks probe tests skippable for bug fixes, and directs config-file reads
  to locate the basename instead of guessing its extension.

v1.65.0

What's Changed

Added

• pidx --report lifecycle [--session ID]. A per-turn ledger — input, cache-read,
  output, calls, result bytes — alongside four cross-turn waste detectors: dupCall,
  resendAfterBlock (a duplicate edit whose earlier attempt was blocked), probeRepeat
  (one target, three turns) and proseOnly (an output-heavy tool-free turn). Waste is
  priced against the session's mean prefix, and the model is printed with the numbers so
  an estimate is never read as an attribution. The detector core is pure, so the case
  corpus pins the model without a session directory.
• SS4 search rule. rg -r / --replace is display substitution, not recursion —
  rg recurses by default, so the flag silently changes what the command reports. Warned
  under the new searchReplaceFlag key.

Changed

• One read budget. SYSTEM.md drops the 500-line window, the ~10 KB whole-file
  bypass and the ~100-line tier for a single rule: one window is one unit inside the
  ~40-line / 2 KB model-visible budget, a larger unit is two deliberate windows, and
  several units are several calls in one parallel block.
• Anchor blocks carry their remedy. A not-found anchor returns the closest region as
  paste-ready oldText on the fragment path, or a bounded window (15 lines / 1.5 KB) plus
  a piread {path, offset, limit} call when no fragment matches — so a block costs no
  remedy round instead of a read plus a full payload re-send.
• --max no longer hides the live session. select_sessions folds a dropped session
  inside the live window back into the slice.
• Post-block discipline in SYSTEM.md: a block concerns one span, so only that region
  is re-sent.

v1.64.0

What's Changed

Added

• skills/self-learn and bash-env nvm loading. scripts/env.sh is a BASH_ENV target that
  loads nvm once for daemon-spawned non-interactive shells (shells whose node already resolves
  skip the ~200 ms source), pinned by tests/unit/pibash-env.test.mjs; settings.json re-enables
  pidx.
• Warn-rule corpus cases. errorSwallow, readerBypass and viewCredit join the gate case
  corpus, with matching updates to anchor-guard, bash-cap, read-coverage and search-guard, the
  case harness, region tests and the plan doc.

Changed

• Rule severities have one home: extensions/permission-gate/config.json. config.ts is
  renamed loader.ts and keeps only mechanics — the rule-key registry, severity coercion, env
  overrides and the hard-rule floor — while the WARN_BY_DEFAULT/OFF_BY_DEFAULT shadow tables
  are deleted. A key the file omits or garbles now fails closed to error with a stderr warning
  instead of consulting a code table; shipped behavior is unchanged because the tracked file lists
  every key. The _comment documents the off/warn/error vocabulary, PGATE_<KEY> one-off
  overrides and the destructiveConfirm/commitMessageLint floor. Measured: gate-config 28/28,
  full check 249/249; loader 300 → 269 lines.

v1.63.0

What's Changed

Added

• pidx --report usage and CONTRIBUTING.md — the usage report answers per-model
  spend, cost tail and prefix trend from pidx's own extract cache (1.73 s → 0.19 s), and
  skills/harness-engineer/scripts/usage-metrics.py is deleted; its last job was walking
  110 MB of session JSONL to derive what the cache already holds. CONTRIBUTING.md states
  time-to-first-PR, the daily loop, conventions and where things live.
• TurnState (extensions/permission-gate/rules/turn-state.ts). The temporal rules (R10, EB,
  LG1, R11, PH1, CN) shared thirteen let/Map bindings scattered through the extension factory,
  with mutationSeq declared below the helper that read it. They now share one value and every
  transition is a function of it. index.ts 1570 → 1501 lines, block reasons unchanged.
• pidx classifies gate blocks. results[].blk carries the block family (anchor, tokens,
  loop, …), matched only at the start of the result text so a session reading the gate's own
  source is not counted as blocked. Extract schema v:2; the cache store version moves to 3 so
  stale caches rebuild.
• tests/tools/audit/gate-metrics.mjs — repeats, failure-streak calls, blocks by family and
  gate token share, computed from a pidx extract instead of one-off scripts.
• tests/gate/region-warn.test.mjs — pins the AG2 severity contract: warn passes, while a
  never-read path and a byte-mismatched anchor still block.
• Tooling rule in AGENTS.md and SYSTEM.md: reuse an existing native tool before writing a
  script, plus the bar a new compiled tool has to clear.
• rust-skills: four compiler-optimization rules from the "Achieving warp speed with Rust"
  guide — opt-flat-storage, opt-registers, opt-loop-unroll, opt-ilp — plus
  assert-hoisting guidance in opt-bounds-check and runtime-share guidance in
  perf-profile-first. 265 → 269 rules.
• npm run audit:session and npm run audit:bin-diff — two reusable audit drivers.
  session-blocks.mjs parses a session JSONL directly and reports chaining, the block ledger with
  gate families and recovery, and re-derives the piread compaction contract from recorded results;
  it caught the compaction drift on real session data. bin-diff.mjs is the byte-for-byte
  differential between two builds of one native tool — shape corpus × argument grid plus a --serve
  stream mode, exit 1 on any difference; it gated the piread scanner refactor.
• npm run bench:piread, bench:pibash and bench:piedit — all three micro-benches already existed in
  tests/tools/bench/ with no entry point and the same stale ../../extensions import, so none had
  run in weeks. All are wired, repaired and listed in tests/README.md.

Changed

• piread scans without per-line reads. Segmenter searches each fill buffer with memchr
  instead of calling read_until per line, and read_text stops decoding and buffering selected
  lines once the head buffer is saturated — they can never be rendered, and the only state those
  calls still moved was a counter nothing reads after the scan. Output is byte-identical (128
  corpus × argument cases, CLI and daemon); a 250k-line / 5.7 MB read goes from 15.7 ms to 7.9 ms
  median, while a 2,500-line read is unchanged.
• piedit borrows its fuzzy view. normalizeForFuzzyMatch ran NFKC over the whole file because
  duplicate counting needs the view even for an exact edit. The view is now borrowed when the
  transformation is provably the identity — ASCII content with no line ending in whitespace — and
  the expect that assumed a materialized view became a fallback to the content itself. Worth
  ~0.5 ms on 589 KB (below bench noise), but it drops a full-content pass and a panic path from the
  request handler.
• AG2 (editRegionGuard) honours its configured severity. Its block bypassed enforce, so
  warn blocked exactly like error. Each anchor decision now carries the rule that owns it and
  severity is applied per rule. The shipped default stays error; plan §12 records the watch
  criteria and flip procedure for relaxing it.
• config.json destructiveConfirm set to error (the hard-rule clamp already forced it).
• custom-tools/ workspace renamed to rust-tools/. Docs, scripts, extensions and tests
  updated; docs/custom-tools-architecture.md → docs/rust-tools-architecture.md; all five
  binaries rebuilt and reinstalled from the new path.
• pre-push-review skill renamed to code-reviewer (/skill:code-reviewer), with
  company-specific context removed from SKILL.md and references.

Fixed

• Compaction tiles every window exactly. compact counted lines with split_lines, which
  drops the body-final empty segment, so a trailing-blank window claimed one line fewer than it
  rendered and head+gap+tail missed it. It now splits on '\n' and attributes each omission from
  max(kept, rendered); renderedCoverageFromBlocks / renderedSpansFromBlocks take the read
  window's end, so the backward anchor maps the final run onto it. Found by the cross-layer
  differential (tests/parity/piread-compaction-walk.test.mjs: real bin/piread output → TS walk),
  locked by the 300-case reconstructs_the_window_exactly sweep and
  trailing_empty_window_line_is_fully_accounted.
• A user-configured AG2 warn had no effect: the anchor guard returned its block directly instead
  of through severity enforcement.
• pibash scanned each output chunk three times. OutputAccumulator::append_decoded_text
  walked a chunk to record newline offsets, walked it again with matches('\n').count(), then a
  third time for rfind. The count now comes out of the first loop. Output is byte-identical
  (15-file cat corpus plus seq 50000, path/timing fields excluded), a 290 KB / 50k-line command
  drops from 12.71 to 11.84 ms median (~7%), and seq 2500 is unchanged. Every pibash benchmark
  class is now below its built-in oracle, so the long-standing seq 50000 acceptance note is
  retired.
• skills/rust-skills never loaded. Its frontmatter had a quoted scalar followed by trailing
  text (- "Achieving warp speed with Rust" (jFransham, gist)), which is invalid YAML, so the loader
  dropped the whole skill: no /skill:rust-skills entry and nothing injected on invocation. The
  scalar is plain now, and tests/unit/skills-frontmatter.test.mjs fails on any loader diagnostic
  or any skills/*/SKILL.md directory that does not load.
• pidx no longer rewrites an up-to-date extract file. The indexer already skipped
  rewriting its store when nothing changed, but still rewrote the extract file that is a pure
  function of that store — 15 MB per warm call on this corpus. A warm run now leaves an
  up-to-date file untouched (444 → 28 ms median, ~16x; full re-parse unchanged at ~2.1 s), the
  output stays byte-identical, and a missing or truncated file is still rebuilt. The store's
  header is also serialized in sorted key order now: two identical cold runs used to write
  different bytes because a HashMap's iteration order leaked into the file.
• docheck resolves --help probes across cores. Every (binary, flag) pair spawns the tool
  once — git's help path drags man and its formatter along — so 2,771 child processes ran
  serially. The probe loop was already isolated from finding emission, so it now fans out over
  the available cores with indexed results. Output byte-identical, npm run doc:check
  3.0 → 1.9 s on four cores.
• benchmark/runs/ is no longer gitignored. Run records are cited from docs/native-tools.md
  and docs/rust-tools-architecture.md, so ignoring the directory broke every link to a committed
  record. The scaffold test now asserts the directory stays trackable instead of asserting the
  ignore line.

Removed

• skills/llm-format-checker. Its rule set duplicated the AGENTS.md docs style that
  skill-manager already enforces; the skill was deleted, its row dropped from
  docs/skills.md, and its auto-apply instruction replaced with the style rule itself.

Docs

• docs/rust-tools-optimization.md — the repeatable path for optimizing a native tool: model
  the cost by stage, freeze the old binary as an oracle, pin the seam before swapping it, run the
  verification ladder, measure with interleaved A/B next to a published noise floor. Worked
  examples carry measured deltas, alongside the traps already paid for (stale daemons, compaction
  budget vs the gate cap, state that goes dead after the scan).
• Plan §10 records measured handler cost (96 µs mean per tool call), §11 the session-log
  measurement (223 sessions, 16,249 calls, 3.7% blocked, AG2 owning 42% of anchor blocks), §12 the
  AG2 watch decision.
• Validation table corrected: stale plan: line references (+6 offset), V6's stated reason (a
  9-byte literal against a 10-byte window is a dead comparison, not a compile error), V7's line,
  V14's evidence, and the absent V12 row. §8's "no binary input channel" narrowed to "no channel
  for compressed text", with binary-media billing and prompt caching added.
• Repeat-call metric corrected from 399 to 70 after the pidx param-hash recount; the earlier
  signature-based count conflated separate edits to the same file.

v1.61.0

What's Changed

Added

• Anchor Guard repeat-miss hint. The 2nd+ anchor block on one file within a user turn appends a loop-break hint (piread the exact target lines, then resend once) to the block reason; count resets per user turn and on compaction (shares LG1's reset). Motivated by audit v2: anchor blocks are the top LLM error class (94/wk) and churn turns (>=4 same-path edits) hold 27% of all blocks+errors. Measured: gate 289/289 (1 new case); audit v2 basis 397 user-turns, 132 blocks, mistake-turns avg 12.7K out-tok vs 2.8K clean.
• scripts/session-audit.py — the turn-by-turn, byte-level session audit (v2) moved into the repo from /tmp: per-user-turn ledger of in/omitted bytes, elision + spill, usage tokens and cost, gate blocks by family, re-fetched read lines, duplicate commands, churn; full JSON to .cache/session-audit.json. Stdlib-only python3; fd --newer <date> scoping. Measured: week rerun from the repo reproduces the report (64 sessions, 397 turns, $62.73 basis).

v1.60.0

What's Changed

Added

• R1x (reReadConsume) revived as the partial-overlap re-read rule. Replaces the old inert self-expiring-dedup semantics (never enabled, zero resend-loops in a week of sessions). warn (new default) appends a non-blocking note when ≥50% of a read's rendered lines were already in context and the file is stat-fresh; error blocks at call time on the requested window and also catches same-batch sibling reads of the same path (the batch-slip shape). Threshold from the 2026-10-26 session audit. Measured: 288/288 gate cases + 28/28 gate-config (6 new); audit basis 149 ≥50%-overlap re-reads/wk (~2.6K re-fetched lines) vs 94 at 70%, 35 R1 catches, 6 same-batch dups across 33 sessions.

v1.59.0

What's Changed

Added

• LG1 loop guard — action-side stop conditions (1b). Evidence-first redesign: a week of this repo's sessions showed the grind is same-target churn (worst: task.md ×9, index.ts ×8 within single user-turns) while consecutive-error streaks never exceeded 3 — so error-streak thresholds would never have fired. LG1 notes (non-blocking, rides the tool_result) at the 4th landed same-path edit/write in a user turn; the 7th arms a one-shot block of that path's next mutation ("summarize what you tried, then ask the user"); 3 consecutive tool errors note once. Counters reset per user turn and on compaction; loopGuard is config-switchable like every rule. Measured: permission-gate 284/284 PASS (4 new LG1 cases), gate-config 26/26 PASS, typecheck + agents-lint + doc-check clean; rule count 22 → 23 (README map corrected).
• Circling-check behavior rule. AGENTS.md Workflow & TDD gains the behavior-space mirror: same file edited 3+ times or the same failing fix retried twice in one turn → stop, state what remains and what failed, ask before continuing. The gate arrests a loop mid-spin; this rule teaches the finish the gate cannot. Measured: AGENTS.md 4,703 → 4,868 bytes; agents-lint 34 bullets, 0 dup/contradictions; doc-check 4/4 clean.

Changed

• Dedup + shrink pass on AGENTS.md + SYSTEM.md. Removed intra-file duplicates (header restated the section list; TDD bullet restated itself; the two Completion bullets overlapped), cross-file duplication (tool list and .nvmrc chain lived in both headers — SYSTEM.md is their one home), and filler wording ("advanced", "high-speed", restated ask-first clauses). Measured: AGENTS.md 4,868 → 4,349 bytes, SYSTEM.md 3,630 → 3,500; agents-lint 33 bullets, 0 dup/contradictions; doc-check 4/4 clean.