harnex 0.12.0 → 0.29.1

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 3ea3eab047cc2e9f086fba4cfb98c6f585ffe888fdee0fb018c35515fe6257b4
-  data.tar.gz: ba40b96028fd01e6ade7517374e7d26d92a5d147a7be2ec08a581cd0203dae62
+  metadata.gz: e36a59564be22ba55cda2815e3e72b59b887bb2a4e0417be776295ef1cb10b82
+  data.tar.gz: 0a9b1355a5bc153e4d6fdd6ef79df81f022c5b6a75f97de0a8f91ef52353e526
 SHA512:
-  metadata.gz: e1e85c57df13dd411b10308659b72376f9491f1c6e4d7c4e5d906f9f78096dc0f82a0ac866c01e356d75cedb5f285b71970d0ea175e5935ca8ed0e3bcf40dd88
-  data.tar.gz: 1283d788ceaff2e608f7e13b7f4f640d65d9f683e35cbe649cc9dd7d0223e094740df146fbe10ed6fa15685bf27671e9059c6378bb2bc8acc5c850aef5f6c9ab
+  metadata.gz: 1794b1c8190bf40109732a821470d9a24772be9e68e46b9c49732ac46bd24fb73a88d07d0097dece3ec43e3b164932903d69ffda7fa6b4f18b244989990c8a52
+  data.tar.gz: 8163c5eb86dacb6174868459fcc9bba2fb87b66c503e61a96f38baefbe2ca93423df3a84fd07da626927402fe5432f4f541eec202a7b4cc4fde131e1b9ff7de4

data/CHANGELOG.md

@@ -1,5 +1,621 @@
 # Changelog
 
+## [0.29.1] - 2026-10-23 | 10:31 AM | IST
+
+Local-only fix release for #88: lanes in repos with a phase allowlist.
+
+### Fixed
+
+- In a repo whose `.harnex/config.json` phase allowlist has
+  `"policy": "reject"`, harnex refused every `delegate`, `ask --write` and
+  `reply` (`meta.phase (none) is not allowlisted`). The run behind a lane
+  is now tagged with phase `delegate` or `ask`. These are harnex's own
+  phases, like `preflight`, so every allowlist accepts them.
+- A lane that harnex refused before the worker started left a `failed`
+  record, so a retry under the same name failed with `lane NAME already
+  exists (failed)` until `discard`. Now a refused new lane leaves no record
+  and no brief. A refused `reply` keeps the state the lane's last turn left.
+  The next `delegate` under the same name also clears a record left this way
+  by 0.29.0.
+
+### Added
+
+- `harnex delegate --phase NAME` and `harnex ask --phase NAME` tag a lane
+  with a work phase such as `implement`. The repo allowlist checks a named
+  phase, and `reply` reuses it.
+
+## [0.29.0] - 2026-10-22 | 06:08 PM | IST
+
+**The agent release.** harnex becomes a tool your coding agents use directly:
+hand a question or a task to a Pi worker (GPT by default) and get the reply
+back as plain text, with a footer saying what happened and what to do next.
+This entry summarizes 0.22.0–0.29.0, the first public release since 0.12.0.
+Website and docs: <https://harnex.zyt.app>. Upgrading:
+`docs/upgrading-to-0.29.md`.
+
+### Highlights
+
+- **Lanes** (0.22):
+  - `harnex delegate [--worktree NAME]` hands a task to a worker, optionally
+    in its own git worktree;
+  - `result` collects parallel lanes;
+  - `reply` follows up with the worker's context (a finished lane resumes its
+    Pi session);
+  - `land` merges and cleans up; `discard` drops a lane; `lanes` shows what
+    each worker is doing right now.
+- **`harnex mcp`** (0.23): the same verbs as native MCP tools for Claude Code,
+  Pi and any MCP client. Long calls report progress, and Claude Code moves them
+  to the background and notifies on completion. `harnex mcp install`
+  registers it for both clients.
+- **The `gpt` subagent for Claude Code** (0.21, MCP relay since 0.23): ships
+  as `agents/gpt.md`, and `harnex mcp install` installs it.
+- **Smoothness work** (0.24–0.28), measured with `evals/agents/run`. Real
+  Claude Code and Pi supervisors, with and without MCP, run seven scenarios;
+  0.28 passed two consecutive full rounds (28/28 each) with no harnex usage
+  errors, no doc lookups and nothing left behind. Along the way:
+  - progress detail;
+  - the hand-merge sweep;
+  - model-failure next steps;
+  - forgiving flags and MCP arguments;
+  - workers never see harnex's own MCP tools.
+
+### Breaking (since 0.12.0)
+
+- Codex adapters, the JSON-RPC transport, `run --fast`/`--legacy-pty` and
+  `preflight --legacy-pty` were removed (0.20); harnex is Pi + GPT first.
+- `harnex ask` waits 9 minutes by default (0.22); use `--timeout none` for no
+  cap.
+- Pi starts with `--no-approve` unless `-- --approve` is passed (0.22).
+
+### Added in 0.29.0
+
+- **Docs:**
+  - `docs/lanes.md`, `docs/mcp.md`, `docs/agents.md` and
+    `docs/upgrading-to-0.29.md`;
+  - `GUIDE.md` now starts with lanes;
+  - `TECHNICAL.md` gains a "Lanes and MCP" architecture section.
+- **The website** `site/`: a monospace, terminal-styled site built from these
+  docs and the live CLI help, deployed as a Holm app to harnex.zyt.app.
+- **The `gpt` agent ships in the gem** (`agents/gpt.md`). `harnex mcp
+  install` installs it at `~/.claude/agents/gpt.md` unless one exists
+  (`--no-agent` skips it). `uninstall` removes only an unmodified copy.
+- **Gem metadata:** a new summary and description, plus documentation and
+  changelog links.
+
+### Changed
+
+- Examples use generic model names (`openai/gpt-5.1`).
+
+## [0.28.0] - 2026-10-22 | 05:15 PM | IST
+
+Owner-authorized local-only release (#87): fixes from round B on 0.27.0.
+
+### Changed
+
+- The MCP `ask` tool description, `harnex ask --help` and the skill index
+  say that an ask worker only reads and cannot run commands; to run tests,
+  scripts or benchmarks, use `delegate`. In round B, a relay sent "run
+  bin/slow-bench" to a read-only ask, and the worker could only read the
+  script.
+- Outside this repo: the `gpt` relay routes a brief that asks to run
+  something or change files to `delegate` (in place), questions and reviews
+  to `ask`, and honours `mode: read`.
+- Evals: `bin/slow-bench` prints a random score and records it in
+  `.git/bench-last-score`. S4 and S6 pass only when the benchmark really
+  ran and the answer quotes that score; a fixed score in the script could
+  be read without running it.
+
+## [0.27.0] - 2026-10-22 | 04:58 PM | IST
+
+Owner-authorized local-only release (#87): fixes from round A on 0.26.0
+(27/28 pass, 0 leftovers, 0 doc lookups).
+
+### Changed
+
+- MCP tools accept the argument names agents reach for from their other
+  tools:
+  - `prompt`, `brief`, `message` or `instructions` for `task`;
+  - `prompt` or `query` for `question`;
+  - `prompt` or `text` for `message`;
+  - `lane`, `id` or `worktree` for `name`;
+  - `name` or `lanes` for `names`.
+
+  A supervisor passed `prompt` to `delegate` and got "task is required".
+- Evals: `--resummarize DIR`, and an `HxErr` column that counts harnex
+  usage friction apart from worker failures harnex reported properly.
+- Outside this repo: the `gpt` agent's description says that `harnex lanes`
+  (or the `lanes` tool) shows what a running worker is doing.
+
+## [0.26.0] - 2026-10-22 | 04:40 PM | IST
+
+Owner-authorized local-only release (#87): fixes from the first full 0.25.0
+eval round (28/28 pass).
+
+### Added
+
+- `harnex land` with no name, and the MCP `land` tool without `names`,
+  land every finished worktree lane the caller started in this repo,
+  oldest first. A supervisor called `land {}` expecting this. `discard`
+  still requires names.
+- `harnex doctor` reports `integrations`: whether the harnex MCP server is
+  registered for Claude Code and Pi, whether the `gpt` agent is linked, and
+  a hint to run `harnex mcp install`.
+- Closed lane records and their briefs are pruned after 30 days; open lanes
+  never are.
+
+### Changed
+
+- README: `harnex mcp uninstall` before `gem uninstall`.
+
+## [0.25.0] - 2026-10-22 | 04:14 PM | IST
+
+Owner-authorized local-only release (#87).
+
+### Fixed
+
+- **Workers never get harnex's own MCP tools.** With harnex MCP registered
+  in Pi, every harnex-launched Pi worker also saw `delegate` and the other
+  harnex tools, and Pi 1.4.0 keeps MCP tools despite a `--tools` allowlist.
+  In an eval, a read-only ask worker used `delegate` to launch a writable
+  lane. Now:
+  - Pi workers start with `--exclude-tools mcp__harnex__…` (merged with the
+    caller's own exclusions; Pi >= 0.99);
+  - allowlisted (read-only) workers start with `--no-mcp` (Pi >= 1.0.4).
+
+### Changed
+
+- Progress lines read `now bash 12s`, without nested parentheses.
+- Lane names generated from a task skip filler words (`pi-slow-bench`, not
+  `pi-run-the-slow-bench`).
+- Evals lift Claude print mode's 10-minute background ceiling
+  (`CLAUDE_CODE_PRINT_BG_WAIT_CEILING_MS=0`), as interactive sessions have
+  none. S6 accepts `tmux capture-pane` as checking on a worker.
+
+## [0.24.0] - 2026-10-22 | 03:51 PM | IST
+
+Owner-authorized local-only release (#87): fixes from the 0.23.0 agent evals.
+
+### Added
+
+- **What a worker is doing now.** The live activity snapshot records the
+  current tool's name and the turn's tool-call count. Tool arguments are
+  still never retained.
+  - MCP progress shows it: `wc: working 2m10s · 7 tool calls · now bash (12s)`.
+  - So does `harnex lanes`: `working (now bash 12s)`.
+  - MCP names ask and delegate lanes up front so progress can follow them.
+- **Hand merges are tidied.** A lane whose branch was merged by hand (into
+  HEAD, worktree clean, at least one commit) is marked landed, and its
+  worktree and branch are removed on the next `delegate`, `result` or
+  `lanes`.
+- **Clearer next steps in the footer.**
+  - Model and provider failures suggest retrying on Pi's default model:
+    `harnex reply NAME "Continue..." --model provider/model` keeps the task
+    through the lane's Pi session.
+  - Lanes with no commits say there is nothing to merge.
+  - Done lanes say `land` merges and cleans up (instead of `git merge`); the
+    MCP instructions say the same.
+- **Evals:**
+  - two new scenarios: S6 (check on a running worker) and S7 (recover from
+    a broken model);
+  - the exit-124 continuation no longer counts as a tool error.
+
+## [0.23.0] - 2026-10-22 | 03:28 PM | IST
+
+Owner-authorized local-only release (#87).
+
+### Added
+
+- `harnex mcp`: a stdio MCP server exposing `ask`, `delegate`, `result`,
+  `reply`, `land`, `discard` and `lanes` as native tools for Claude Code, Pi
+  and any MCP client. Each tool runs the CLI verb and returns its text.
+  - Blocking calls send `notifications/progress` every 15s; the default wait
+    is 25 minutes.
+  - Calls run concurrently.
+  - A cancel stops the wait, never the worker.
+- `harnex mcp install|uninstall|status`: register the server for Claude Code
+  (`claude mcp add --scope user`) and Pi (`~/.pi/agent/mcp.json`, exposure
+  `direct`, 2h request timeout). It registers the RubyGems wrapper, so no
+  version-manager shim or client cwd is involved.
+- `harnex doctor` capabilities add `mcp`.
+- Lane verbs accept habits from `harnex run` instead of refusing them:
+  - `--id` and `--tmux [NAME]` name the lane;
+  - `--auto-stop` and `--wait-done` are accepted and ignored;
+  - `--wait-timeout` is an alias for `--timeout`;
+  - `--context` and `--context-file` give the task;
+  - `reply` takes `--id` and `--message`.
+- `evals/agents/run` routes: `claude-cli` (no MCP, no `gpt` agent) and
+  `pi-cli` (`--no-mcp`) beside the configured `claude` and `pi`.
+
+## [0.22.0] - 2026-10-22 | 03:07 PM | IST
+
+Owner-authorized local-only release (no RubyGems publication); breaking
+changes allowed (#87). Plan: `koder/plans/39_agent_smoothness.md`.
+
+### Added
+
+- **Lanes: a small agent-facing layer over Pi workers.**
+  - `harnex delegate [--worktree NAME] TASK`: one call hands a task to a
+    worker and prints its reply plus a lane footer: state, time, model, cost,
+    commits, diffstat and next commands. `--worktree NAME` creates
+    `.harnex/worktrees/NAME` on branch `harnex/NAME`. `--no-wait` starts
+    parallel lanes.
+  - `harnex result [NAME...]`: wait for lanes and print each result as it
+    lands. With no names it covers every uncollected lane the caller
+    started in this repo.
+  - `harnex reply NAME MESSAGE`: a follow-up. A busy worker takes it
+    mid-turn; a finished lane is relaunched on its own Pi session
+    (`pi --session`), so it keeps its context.
+  - `harnex land NAME`: merge `harnex/NAME` into the current branch, then
+    remove the worktree and branch. Leftovers are committed first; a
+    conflict leaves the merge in progress for the caller.
+  - `harnex discard NAME` and `harnex lanes`.
+  - Lane records live in `STATE_DIR/lanes/`. Blocking commands default to
+    `--timeout 9m`, and exit 124 names `harnex result NAME`.
+- `harnex wait --since TIME` ignores events from an earlier launch of the
+  same id (each launch restarts the events-log seq).
+- `harnex doctor` capabilities add `lanes`.
+- `evals/agents/run`: agent smoothness evals. Real supervisors (Claude Code
+  headless, Pi print mode) run five scenarios against a fixture repo,
+  scored for correctness, tool calls, errors, doc lookups and leftovers.
+
+### Changed
+
+- **Breaking:**
+  - `harnex ask` defaults to `--timeout 9m` (`none` for no cap).
+  - Every ask is a lane, so a timed-out ask names `harnex result ID`, and
+    `harnex reply ID "..."` follows up with the worker's context.
+    `ask --resume ID` still works and no longer needs `--repo`.
+  - Pi starts with `--no-approve` unless the caller passes `--approve`.
+- Top-level `harnex --help` leads with the agent verbs;
+  `harnex agents-guide claude` is rewritten around them.
+- Help examples use Pi session ids (`pi-i-42`) instead of Codex-era ones.
+
+### Fixed
+
+- Harnex state in a repo no longer shows up as untracked work: writing a
+  dispatch row under `.harnex/` adds `.harnex/` to the clone's
+  `info/exclude`. This also unblocks `git worktree remove` for lane
+  worktrees.
+
+## [0.21.0] - 2026-10-22 | 12:40 PM | IST
+
+Owner-authorized local-only release (no RubyGems publication). Makes Pi/GPT
+workers usable as a Claude Code subagent type (#86 slice 1): the `gpt` agent
+(kept in `~/Projects/pi/.claude/agents/gpt.md`) relays briefs through
+`harnex ask` and needs `ask --resume` to outlast the 10-minute Bash cap.
+
+### Added
+
+- `harnex ask --resume ID --repo ROOT [--timeout DUR]` (#86): keep waiting on
+  an ask that timed out and print its answer like the original ask (same
+  exit codes, the full answer even when truncated in the result). A timed-out
+  ask now names this command on stderr instead of a raw `harnex wait`, so a
+  Claude Code subagent loops on one command shape under the 10-minute Bash
+  cap. `--json` output still carries the run's own `resume` field.
+- `harnex doctor` capabilities add `ask_resume`.
+
+## [0.20.0] - 2026-10-20 | 09:55 PM | +04
+
+Owner-authorized local-only release (no RubyGems publication). Harnex is now
+Pi + GPT only: Codex support is removed (owner decision 2026-10-20: "we are
+going to be exclusively pi & gpt models").
+
+### Removed
+
+- The `codex` (PTY) and `codex app-server` (JSON-RPC) adapters, the
+  `stdio_jsonrpc` session transport, the app-server client, and their
+  fixtures and tests. `harnex run codex` is no longer special-cased (it would
+  wrap a `codex` binary as a generic PTY CLI).
+- `harnex run --fast` and `--legacy-pty`; Codex service-tier forcing;
+  `harnex preflight --legacy-pty`; `docs/codex-appserver.md`.
+- `harnex doctor` Codex checks: it now checks Pi by default and only Pi
+  (`--adapter pi|all`).
+- Issues #40, #42 and #43 (Codex recovery/fallback/throughput work) closed.
+
+### Changed
+
+- Shared session behavior previously tested through Codex (receipts and
+  claims, completion-notification ordering, boot-failure classification,
+  stop sequencing, auto-stop, status fields, usage pricing, `preflight`
+  end to end) is now tested through Pi or adapter-neutral fakes.
+- Dispatch-row schema fields are unchanged; `agent.service_tier`,
+  `actual.rate_limits` and Codex error details stay `null` for new rows.
+- `harnex doctor` capabilities add `ask`, `final_message` and `owner_tag`.
+- Docs, guides, recipes and repo instructions are Pi-only; session ids use
+  `pi-<role>-<tag>`.
+
+## [0.19.0] - 2026-10-20 | 06:35 PM | +04
+
+Owner-authorized local-only release (no RubyGems publication). Makes Pi
+(GPT-backed) workers usable as native tools from Claude Code sessions and
+subagents (#85), driven by live Claude-subagent stress runs.
+
+### Added
+
+- `harnex ask "QUESTION"`: one blocking call that hands a question to a Pi
+  worker (read-only tools by default; `--write` to allow edits) and prints its
+  answer. `--model`, `--fallback-model` (one more try on provider capacity),
+  `--effort`, `--file/--section`, `--timeout`, `--json`; stderr names the
+  lane, time and model. A read-only ask records in the global stream, so the
+  repository stays untouched.
+- Turn results carry the worker's reply: `final_message` (at most 4,000
+  characters), `final_message_path` (full text), `final_message_truncated`,
+  plus the lane's `end_sha`, `commits` (since launch), `turn_commits` and
+  `model`, in events, `wait`, `send --wait-done` and owner-stream lines.
+- `harnex run --wait-done [--wait-timeout DUR]`: launch and block until the
+  first turn's result in one call; on 124 the output names the `resume`
+  command.
+- `--owner-tag TAG` on `run` and `watch --owner`: a subagent's own owner
+  (`<session>/<TAG>`), so its stream holds only its lanes and no backlog.
+- `watch --owner --match GLOB` filters lanes by id; `--until-idle` blocks
+  until no matching lane is running and prints every result on the way.
+- `send --stall-after DUR` (with `--wait-done`); `send --wait-done` timeouts
+  print `resume`.
+- Status and stall lines name Pi's running shell commands
+  (`running_commands`; "stalled … while running `sleep 150`").
+- `harnex agents-guide claude`: the operating guide for Claude Code.
+
+### Changed
+
+- `watch --owner --once` prints every event ready in its pass, not only the
+  first.
+- The owner stream delivers each turn's result once and in settle order, even
+  when several turns end between polls (owner records keep a bounded
+  per-turn history). Previously a turn replaced by the next before a poll was
+  lost.
+- `harnex send` to an exited session fails fast (exit 1, `status: "exited"`)
+  instead of spending its whole timeout on lookup.
+- `send --wait-done` output: `status` is the turn's verdict and
+  `delivery_status` the delivery's.
+- `wait --until done` on a session already gone reports `process_state:
+  "exited"` and names a caller's stop `stopped_by_caller`.
+- Stall lines carry `turn` and `elapsed_s`; durations under 10 minutes print
+  as `3m05s`.
+- Durations accept compound forms (`8m30s`, `1h30m`); `--context-section`
+  accepts a heading with its marks (`"## Task"`).
+- A follow-up whose run settled while the send was in flight is sent as a
+  plain prompt, and follow-ups are counted for the turn they joined.
+- `resume` hints name the checkout's `bin/harnex` when running from source.
+
+## [0.18.0] - 2026-10-20 | 04:41 PM | +04
+
+Owner-authorized local-only release (no RubyGems publication). Closes the
+subagent-orchestration gaps found by Holm Queues 159/160 (#79-#84).
+
+### Added
+
+- `harnex run --context-file PATH [--context-section TEXT]` sends a brief from
+  disk, or only the part after a heading (to the next heading of its level)
+  or a plain marker line (to end of file), and refuses an empty result.
+  `--meta-file PATH` reads `--meta` JSON from a file (inline `--meta` keys
+  win). The brief path and SHA-256 of the sent text land in
+  `meta.context_file` / `meta.context_sha256` (#83).
+- `harnex watch --owner --once [--timeout DUR] [--cursor NAME]`: print the
+  first undelivered result or stall, advance the cursor, exit 0 (124 on
+  timeout). One blocking call covers every lane an owner launched; named
+  cursors keep a parent and its subagent apart (#84).
+- Per-turn completion (#81): `harnex send` reports `after_seq`;
+  `harnex wait --until done --after-seq N` waits for the turn after that point;
+  `harnex send --wait-done` does both in one call. `task_complete` /
+  `task_failed` events carry `turn`, and the owner stream reports every turn's
+  result (a nudged turn included) instead of one per session.
+- `provider_capacity` outcome class: a Pi turn that fails after Pi's own
+  provider retries ran out (`auto_retry_end success:false`). In status JSON,
+  dispatch rows, the owner stream and `orchestration report`; fatal by default
+  for `--failure-budget`, as `task_failed` was (#82).
+- Turn results carry `follow_ups: {sent, pending}` when follow-ups joined the
+  turn (from Pi's `queue_update` counts; `pending: 0` means each reached the
+  model), and owner events add `settled_at`. A later turn's `elapsed_s` runs
+  from that turn's start. `harnex run --tmux` prints `tmux_target`.
+
+### Changed
+
+- A send to a busy Pi session is handed to Pi at once as a follow-up
+  (`delivery: "follow_up"`), delivered when the current turn ends, instead of
+  waiting in the 120 s inbox and expiring unread. Other adapters: an expired
+  message makes `send` exit 1 with `status: "expired"`, and a send that times
+  out withdraws its queued message before exiting 124 (#79).
+- Worktree lanes (`--cwd` in a different Git toplevel than `--root`) are
+  observed in the lane: `commit_sha`, branch, receipt, activity and acceptance
+  describe the lane; the end row adds `root_commit_sha` (#80).
+- `run --tmux --context-file` hands the brief to the inner runner by path,
+  so long briefs no longer hit tmux's command-length limit.
+
+## [0.17.0] - 2026-10-18 | 01:13 PM | +04
+
+Owner-authorized local-only release (no RubyGems publication).
+
+### Added
+
+- `harnex watch --owner [ID]`: one owner-scoped stream of compact JSON lines
+  (`harnex.owner_event.v1`, with a readable `msg`) for every dispatch an owner
+  started: `started`, `done`, `rejected`, `failed`, `stalled` (with
+  `--stall-after`), `stopped`, `timeout`. No heartbeats; runs until killed or
+  `--max-wait`. A per-owner cursor delivers each result once across restarts.
+  Built for Claude Code's `Monitor` (#77).
+- `harnex run` from a Claude Code Bash shell records `CLAUDE_CODE_SESSION_ID`
+  as the dispatch owner, as Pi does with `PI_SESSION_ID` (Pi's wins if both
+  are set; explicit `--orchestration-session-id` wins over both). The tmux
+  hand-off marks the owner resolved, so a tmux server's stale environment
+  cannot re-attribute a dispatch.
+- Documented the stable `harnex status --json` fields for status-line and
+  panel consumers.
+
+## [0.16.1] - 2026-10-18 | 12:41 PM | +04
+
+Owner-authorized local-only release (no RubyGems publication).
+
+### Added
+
+- `harnex watch --stall-after DUR` (and `wait --until done --stall-after DUR`)
+  exits 4 (`wait_result: stalled`) once the live session shows no progress for
+  DUR, leaving the worker running. Pi and other structured adapters use the
+  most recent turn-start/model/tool event during an active turn
+  (`clock: activity`); PTY adapters fall back to transcript idle
+  (`clock: output_log`). Stalls write no marker and skip
+  `--stop-on-terminal`. The predicate (`Harnex::Stall`) is shared for #77
+  (#76).
+
+## [0.16.0] - 2026-10-18 | 12:55 AM | +04
+
+Owner-authorized local-only release. Owner decision (2026-10-18): Harnex is
+installed locally and is not published to RubyGems.
+
+### Added
+
+- Final gates: `harnex run --final-gate '<cmd>'` (repeatable; or meta
+  `final_gates`) declares final acceptance commands. Receipts gain
+  `validation.commands_had_failures` (whole-history, harness-observed) and
+  `validation.final_gates` (`declared`, `observed_pass`, `observed_fail`,
+  `missing`, `status: pass|fail|unknown|undeclared`, per-gate last exit).
+  With gates declared, `validation.status` reports them, so an exploratory or
+  intermediate failure no longer marks a green final run as `fail`. Matching
+  is exact, after unwrapping a single `bash -lc`/`sh -c` transport wrapper.
+  Without gates, behavior is unchanged. Diagnostics only: acceptance and
+  outcome classes do not change (#68).
+- Dispatch-row `outcome.class` gains process classes from signals Harnex
+  already holds: `boot_failure`, `permission_prompt`, `timeout`,
+  `stopped_by_caller`, `process_failed`. They refine an `unknown`/`task_failed`
+  terminal or the receipt rejection of a process that never signalled
+  completion; completion and proof verdicts are never overridden, and the
+  vocabulary is closed with an `unknown` fallback.
+- `harnex orchestration report` counts worker outcome classes per raw attempt
+  and per retry family (joined by parent attempt/dispatch id), accepts
+  `--queue-id` as well as `--run-id`, and takes an optional
+  `--failure-budget N [--budget-unit attempts|families] [--fatal-classes A,B]`
+  that adds a `failure_budget` verdict and exits 3 on breach (#57).
+- `harnex preflight ADAPTER [--cwd DIR] [--commit-check]` runs one tiny real
+  dispatch through `harnex run` and prints a JSON verdict
+  (`harnex.preflight.v1`): registration, boot, permission prompt, task read
+  from the task root (default `/tmp`), workspace write, harness-observed
+  command exit (via a final gate), receipt, worker claims, completion, and
+  stop, plus a scratch-repo commit with `--commit-check`. Each failure is a
+  distinct check and failure class; a turn that fails before any work reports
+  its (redacted) error as the root cause. Usage, cost and context telemetry
+  provenance is reported per adapter. Exits non-zero on any failed check. Its
+  dispatch row carries `phase=preflight` (exempt from repo phase allowlists)
+  and `orchestration report` excludes such rows unless `--include-preflight`.
+  Workers are stopped through the session API and carry a `--max-runtime`
+  backstop, so a killed or timed-out preflight cannot orphan them (#56).
+- `harnex doctor` capabilities add `final_gates: 1`, `outcome_classes: 1` and
+  `preflight: 1`.
+
+### Changed
+
+- `harnex status` (default current-repo view) now says how many sessions are
+  active in other repos, with up to three ids, instead of a bare
+  "No active harnex sessions" that read as global emptiness when workers run
+  in sibling worktrees. `--repo`, `--all` and `--json` output is unchanged
+  (#26).
+- A bare `harnex wait` (process-exit fence) prints a stderr `note:` when the
+  session was not started with `--auto-stop`, and again if its turn settles
+  while the wait is blocked, pointing at `--until done`. Events from an
+  earlier run of the same id are ignored. Status JSON gains `auto_stop` (#49).
+- Events JSONL `ts` now has millisecond precision (`iso8601(3)`), so teardown
+  races can be measured from the events log (#36/#37 follow-up). Readers that
+  parse ISO-8601 are unaffected; older rows keep whole seconds.
+
+### Fixed
+
+- Two status tests that sat below a `private` keyword never ran; they run now.
+- Test-suite flakes: cross-repo status hint tests no longer see registries
+  leaked by other suites, and the detached on-done smoke waits for teardown.
+
+## [0.15.0] - 2026-10-16 | 07:20 AM | +04
+
+Owner-authorized local-only release; public publication, tag and push remain deferred.
+
+### Added
+
+- Pi RPC receipts now attest built-in `bash` commands: bounded command text,
+  integer exit from `result.structuredContent.exit_code`, status and duration,
+  correlated by `toolCallId`. Output is never persisted. Pi bash calls count in
+  `commands_executed`; other tools stay `tool_calls` (#70).
+- Receipts gain `command_observation: "partial"` and
+  `observed.commands_unobserved`: Pi command calls without structured exit
+  evidence (abort, timeout, extension override, older Pi, in flight at
+  teardown) are counted rather than inferred, and `validation.status` reads
+  `unknown` instead of `pass` while any exist. Live verified on Pi 1.2.0/1.3.0.
+- `harnex doctor` advertises `capabilities.pi_command_proof: 1`; consumers
+  should check it before treating Pi receipt commands as proof.
+
+### Fixed
+
+- `wait`/`watch` observers no longer exit 1 when a sibling session's registry
+  file is removed between the session scan's glob and read (stop/cleanup
+  race); `doctor`'s registry sweep is guarded the same way (#75).
+
+## [0.14.1] - 2026-10-05 | 09:55 PM | +04
+
+Owner-authorized local-only release; public publication, tag and push remain deferred.
+
+### Fixed
+
+- New dispatch records and reconciliation repair appends store paths under the
+  user's home as `~`/`~/...`, without mutating runtime records or rewriting
+  existing rows. Nested data and home-prefix boundaries are preserved (#74).
+- Canonical reconciliation accepts equivalent absolute and home-relative paths
+  while retaining genuine payload conflicts. Status results keep their absolute
+  repository roots, and retention still protects portable live-record paths.
+
+## [0.14.0] - 2026-09-28 | 08:25 PM | +04
+
+Owner-authorized local-only release; public publication, tag and push remain deferred.
+
+### Fixed
+
+- Explicit idle Pi shutdown preserves the latest settled accepted/no-change
+  proof, including follow-up turns. Busy/newly dispatched work cannot reuse an
+  older receipt; unexpected transport loss and malformed output still fail.
+  Raw child exits remain observable separately from logical cleanup status (#69).
+- `wait`/`watch` observer deadlines now include slow registry, event, HTTP and
+  exit-grace probes. A private observer process enforces the monotonic cap without
+  stopping the worker; late results cannot override timeout (#71).
+
+### Added
+
+- Opt-in `wait/watch --heartbeat DUR` streams flushed diagnostics on stderr while
+  preserving one final JSON stdout result. `wait --max-wait` aliases `--timeout`.
+- Structured work/model/tool activity clocks distinguish real events from UI/log
+  freshness, with unknown PTY clocks and no retained thinking-stream content.
+- Bounded stop reason/origin survives events, status, receipts and dispatch-end
+  telemetry. Expected cancellation is distinct from provider disconnection.
+- `run --max-runtime DUR` enforces a fixed runner-owned worker budget independent
+  of supervisors, prompts, retries or UI chatter. Active expiry fails work; idle
+  cleanup preserves accepted proof. Doctor exposes versioned capabilities (#71).
+
+## [0.13.0] - 2026-09-23 | 01:56 AM | +04
+
+Local-only activation approved by the owner. Public RubyGems publication, Git
+tag, and push are deferred; later publication must use the exact verified gem
+built from this commit rather than rebuilding version 0.13.0.
+
+### Added
+
+- Dispatch telemetry now has an opt-in local archive bridge. A stable sidecar
+  lock encloses archive preparation and append, preventing atomic rotation from
+  stranding writers on an old inode. A separate complete-history reader feeds
+  history, canonical reconciliation, latest-row lookup, status, and attempt
+  accounting. Helper failures never drop a new dispatch row, while unavailable
+  or malformed complete history fails explicitly instead of appearing empty.
+- `harnex telemetry archive-capability --json` exposes the local helper protocol
+  version and exact archiver/history environment names so automation can refuse
+  to trim active dispatch history against an older installed Harnex.
+
+## [0.12.1] - 2026-09-19 | 12:30 PM | +04
+
+### Fixed
+
+- Pi-launched dispatches now capture the invoking primary's `PI_SESSION_ID`
+  before crossing detached/tmux process boundaries and publish a durable,
+  owner-addressed, per-attempt completion record. Reusing a dispatch name no
+  longer erases an unacknowledged attempt, and nested workers never implicitly
+  nominate themselves as completion consumers.
+- Addressed delivery registration fails before the initial structured prompt
+  when it cannot be persisted. A terminal delivery-write failure leaves the
+  registered running record available for native `harnex watch` recovery.
+  Existing typed markers and optional `--on-done` hooks remain compatible.
+
 ## [0.12.0] - 2026-09-03 | 10:09 AM | IST
 
 ### Added