hubkit-cli 0.1.0__tar.gz → 0.2.0__tar.gz

{hubkit_cli-0.1.0 → hubkit_cli-0.2.0}/CHANGELOG.md

@@ -4,8 +4,27 @@ All notable changes to this project are recorded here. The format follows [Keep
 
 ## [Unreleased]
 
+## [0.2.0] - 2026-11-06
+
+### Added
+- Commands: `hubkit digest [--since]`, `hubkit trust record|ls`, `hubkit inbound ingest`, `hubkit work verify`, `learn.py address --check CMD`.
+- The warn-only PreToolUse hook (`hubkit hooks --merge-guard`, matcher `Bash|TaskStop|AskUserQuestion`) now also warns on posts without a typed yes, banned commands, repeated commands, a whole-suite `pytest`, `TaskStop` without a stop word, and a decision card whose first option is not marked (Recommended). The routed hook warns on a brief with no done-when, output or estimate line, on a brief that forbids the fix a blocker needs, and when more than `routed.max_agents` agents hold work.
+- Stop gate reasons: a "tests pass" or "it works" claim in a turn that ran no command, a PR or issue number with no link, an approval while CI at the head is not green, built work not pushed, an ask asked 3 times with no answer, and an unanswered prompt not re-read before "done".
+- Order close: evidence names in a kind's `sha_evidence` must hold a commit sha.
+- Watcher: stale finished agents (`watcher.stale_agent_min`), replies owed to other people (`watcher.reply_owed_min`), and own work not verified for `watcher.verify_stale_min`.
+- `doctor` warns on a hubkit hook registered twice in `settings.json` and on no scheduled `hubkit eod`.
+- Learning loop: `address --check`, an overlay cap of 200 lines, expiry and prose-only lists in `review`, and an ineffective mark on a class that repeats after it was addressed.
+- New tests (336 pass on Linux) and revert entries for each guard.
+
+### Changed
+- Correction patterns in `intents.example.toml` match insults aimed at the hub and no longer fire on one aimed at a tool.
+- A routed item no longer closes on the peer's own "done"; it closes on an artifact (merged or closed PR, or a moved head with CI not failing). A merged or closed PR closes a review item.
+- A one-line resume command may be given in the message as typed (`SKILL.md`).
+
+## [0.1.0]
+
 ### Added
-- Install like miseenplace: `uv tool install hubkit-cli` (wheel built and installed in a scratch venv; not published), `hubkit skills --install DIR` (hub, hub-mep, hubkit-setup), `hubkit hooks [--skills-dir DIR] [--merge-guard]` prints the settings.json hooks block, and a `hubkit-setup` skill.
+- Install like miseenplace: `uv tool install hubkit-cli` (published on PyPI as 0.1.0), `hubkit skills --install DIR` (hub, hub-mep, hubkit-setup), `hubkit hooks [--skills-dir DIR] [--merge-guard]` prints the settings.json hooks block, and a `hubkit-setup` skill.
 - macOS and Windows support through one platform module (`hubcore/compat.py`); only Linux has been run so far.
 - Blockers (`order block|result`), watcher fixes, peer-report closing, `asks ls`, and warn-only merge guard.
 - Apache-2.0 license, `README.md` rewrite, `docs/` (architecture, configuration, commands, troubleshooting), `CONTRIBUTING.md`, `SECURITY.md`.

{hubkit_cli-0.1.0 → hubkit_cli-0.2.0}/CONTRIBUTING.md

@@ -7,7 +7,7 @@ Thanks for looking. hubkit is small on purpose: standard library only, Python 3.
 ```bash
 git clone https://github.com/ankitkpandey1/hubkit && cd hubkit
 pip install pytest ruff mypy coverage
-python3 -m pytest tests -q                       # about 2 minutes: real subprocesses
+python3 -m pytest tests -q                       # about 4 minutes: real subprocesses
 ruff check .
 mypy scripts/hubcore scripts/learn.py scripts/hubkit
 python3 tests/run_reverts.py                     # slow; run it when you touch a guarded behaviour

{hubkit_cli-0.1.0 → hubkit_cli-0.2.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: hubkit-cli
-Version: 0.1.0
+Version: 0.2.0
 Summary: Session supervisor for Claude Code: hooks-based multi-agent orchestration with order clocks, routed-work tracking, a watchdog and a stop gate, standard library only.
 Project-URL: Repository, https://github.com/ankitkpandey1/hubkit
 Project-URL: Documentation, https://github.com/ankitkpandey1/hubkit/tree/main/docs
@@ -45,14 +45,15 @@ In a live hub, the prompt hook opens that order itself from your sentence, and t
 - **Order clock.** Orders open from your own prompts (never from peer or harness text), each with a first-action limit, a stall limit and a definition of done checked by evidence on close.
 - **Routed-work tracking.** A hook on `SendMessage` and `Agent` records what the hub asked of whom, one item per (owner, PR). It follows last heard-from, moved heads and GitHub reviews at the PR head.
 - **Watchdog and wake-up.** `hubkit tick` (cron) and `hubkit watch` (a background process) raise alerts with back-off; a watch exit wakes an idle hub.
-- **Stop gate.** The Stop hook refuses to let the turn end while the hub owes something (no watch running, a stalled item, an unbacked verdict, a question to you that blocks nothing).
-- **Status, end-of-day review.** A status block at session start and on each prompt, a daily archive, and `hubkit eod`.
+- **Stop gate.** The Stop hook refuses to let the turn end while the hub owes something (no watch running, a stalled item, an unbacked verdict, a question to you that blocks nothing, a "tests pass" claim with no command run, a PR number with no link, an approval on red CI, built work not pushed).
+- **Tool guards.** An optional warn-only PreToolUse hook (`hubkit hooks --merge-guard`) warns about an unordered merge, a post without your typed yes, a banned command, stopping an agent and an unmarked decision card. It never denies.
+- **Status, digest, end-of-day review.** A status block at session start and on each prompt, a daily archive, `hubkit digest` for when you come back, `hubkit eod`, a per-class `hubkit trust` ledger, and `hubkit inbound ingest` to track asks from other people so none goes unanswered.
 
 `SKILL.md` is the behaviour contract the hub session reads. This file is the operator view.
 
 ## Requirements
 
-- Linux. hubkit has only been run on Linux; macOS and Windows code exists but has never been run (see [Platform support](#platform-support)).
+- Linux. hubkit has only been run on Linux so far (see [Platform support](#platform-support)).
 - Python 3.14 or newer (the code uses `except A, B:` without parentheses, PEP 758). Standard library only.
 - [Claude Code](https://code.claude.com), which provides the hooks and the skills directory.
 - Optional: `gh` (GitHub CLI) on `PATH` for PR checks; without it items show "PR unknown". Optional: [`mep`](https://pypi.org/project/miseenplace-cli/) for a personal task list in the status block; without it the block shows `mep: UNVERIFIED (mep CLI not found ...)`, and `[mep] enabled = false` silences that.
@@ -64,7 +65,7 @@ Install hubkit with `uv tool install hubkit-cli`, run `hubkit skills --install D
 for your tool, then run the hubkit-setup skill and follow it. Ask me only for things you cannot find out yourself.
 ```
 
-hubkit runs on Claude Code hooks, so the skills directory is Claude Code's: `~/.claude/skills` (user level) or `.claude/skills` (project level). The PyPI name `hubkit-cli` is free as of 2026-11-06 and is not published yet; until it is, install straight from GitHub with `uv tool install git+https://github.com/ankitkpandey1/hubkit` (or from a clone with `uv tool install ./hubkit`).
+hubkit runs on Claude Code hooks, so the skills directory is Claude Code's: `~/.claude/skills` (user level) or `.claude/skills` (project level). `hubkit-cli` is on [PyPI](https://pypi.org/project/hubkit-cli/). To install the latest commit instead: `uv tool install git+https://github.com/ankitkpandey1/hubkit`.
 
 `hubkit skills --install DIR` copies three skills into `DIR`: `hub` (the rules and the scripts the hooks run), `hub-mep` (optional) and `hubkit-setup` (finds your facts, writes the config, registers the hooks, binds, proves it with `hubkit doctor`). Run it again after an upgrade to refresh the copies. `hubkit hooks --skills-dir DIR` prints the hooks block for `~/.claude/settings.json`.
 
@@ -140,12 +141,13 @@ Every key is documented in [docs/configuration.md](docs/configuration.md).
 
 | Behaviour | Strength |
 |---|---|
-| The Stop hook holds the turn while the hub owes something: no watch running, a stalled item, an unbacked verdict, an unsolved blocker, an item over its estimate, a "done" while your newest prompt is unread | Enforced (blocks the turn end) |
+| The Stop hook holds the turn while the hub owes something: no watch running, a stalled item, an unbacked verdict, an unsolved blocker, an item over its estimate, a "done" while your newest prompt is unread, a claim of passing tests with no command run, an approval while CI is not green | Enforced (blocks the turn end) |
 | An order closes only with the evidence its kind requires | Enforced (the close is refused) |
-| A learned class counts as fixed only if a rule for it exists | Enforced (the address is refused) |
-| A second job sent to a busy owner while an order is open; a brief that bans the action a blocker needs | Warning only |
-| `gh pr merge` with no live order naming that PR (the merge guard, `hubkit hook pretool`) | Warning only, never denies |
-| `hubkit doctor` warning when your `intents.toml` lacks keys the example has | Warning only |
+| A learned class counts as fixed only if a rule for it exists (and, with `--check`, its command passes) | Enforced (the address is refused) |
+| A merge order closes only with the merge commit sha | Enforced (the close is refused) |
+| A second job sent to a busy owner while an order is open; a brief that bans the action a blocker needs; a brief with no done-when, output or estimate line; more than 4 agents holding work | Warning only |
+| `gh pr merge` with no live order naming that PR, a post without your typed yes, a banned command, `TaskStop`, a repeated command, a whole-suite `pytest`, an unmarked decision card (`hubkit hook pretool`) | Warning only, never denies |
+| `hubkit doctor` warning when your `intents.toml` lacks keys the example has, a hook registered twice, or no scheduled `hubkit eod` | Warning only |
 
 Details and exact rules: [SKILL.md](SKILL.md) and [docs/architecture.md](docs/architecture.md#stop-gate-gatepy).
 
@@ -168,19 +170,11 @@ python3 tests/run_reverts.py        # each guard reverted in a copy must fail it
 
 ## Platform support
 
-All OS-specific code is in one module, `scripts/hubcore/compat.py` (standard library only): process start time and parent (`/proc` on Linux, `ps` on macOS, kernel32 through `ctypes` on Windows), file locks (`flock`, or `msvcrt` on Windows), the scheduler check (`crontab -l`, or `schtasks /query` on Windows) and UTF-8 stdio.
-
-| OS | Status |
-|---|---|
-| Linux | Full test suite run and passing here (ruff, mypy, pytest). |
-| macOS | The `ps` code path is exercised on Linux against the real `ps` (`tests/test_compat.py`), and mypy checks it with `--platform darwin`. Never run on a Mac. Not claimed to work until someone runs it there. |
-| Windows | The `ctypes` and `msvcrt` code is type-checked with `--platform win32` only. Never run on Windows. Tests that need a POSIX `sh` shim for `gh`, `crontab` or a notifier are skipped there. Not claimed to work until someone runs it there. There is no default desktop notification on Windows: set `tools.notify_cmd`. |
-
-There is no CI. Update this table only from a real run on that OS. On Windows the schedule is a Task Scheduler task whose name or command holds `hubkit` (for example `schtasks /create /sc minute /mo 5 /tn hubkit-tick /tr "python %USERPROFILE%\.claude\skills\hub\scripts\hubkit tick --if-bound"`), and hooks in `settings.json` call `python ...\hubkit hook <event>`.
+hubkit has only been run on Linux so far, and the full test suite, ruff and mypy pass there. The code has paths for macOS and Windows, all in `scripts/hubcore/compat.py` (process lookup, file locks, scheduler check, UTF-8 stdio), but they have never been run on those systems. Treat them as unsupported until someone does.
 
 ## Limits and known gaps
 
-- Python 3.14 only. macOS and Windows are written but unproven; see Platform support.
+- Python 3.14 only, and Linux only so far; see Platform support.
 - It targets one hub per machine. A second live hub is refused at bind.
 - Order creation takes a blocking `flock` (`scripts/hubcore/orders.py`); a process that holds the lock forever stalls order opens.
 - A hook payload with no `origin` field is treated as human input.

{hubkit_cli-0.1.0 → hubkit_cli-0.2.0}/README.md

@@ -31,14 +31,15 @@ In a live hub, the prompt hook opens that order itself from your sentence, and t
 - **Order clock.** Orders open from your own prompts (never from peer or harness text), each with a first-action limit, a stall limit and a definition of done checked by evidence on close.
 - **Routed-work tracking.** A hook on `SendMessage` and `Agent` records what the hub asked of whom, one item per (owner, PR). It follows last heard-from, moved heads and GitHub reviews at the PR head.
 - **Watchdog and wake-up.** `hubkit tick` (cron) and `hubkit watch` (a background process) raise alerts with back-off; a watch exit wakes an idle hub.
-- **Stop gate.** The Stop hook refuses to let the turn end while the hub owes something (no watch running, a stalled item, an unbacked verdict, a question to you that blocks nothing).
-- **Status, end-of-day review.** A status block at session start and on each prompt, a daily archive, and `hubkit eod`.
+- **Stop gate.** The Stop hook refuses to let the turn end while the hub owes something (no watch running, a stalled item, an unbacked verdict, a question to you that blocks nothing, a "tests pass" claim with no command run, a PR number with no link, an approval on red CI, built work not pushed).
+- **Tool guards.** An optional warn-only PreToolUse hook (`hubkit hooks --merge-guard`) warns about an unordered merge, a post without your typed yes, a banned command, stopping an agent and an unmarked decision card. It never denies.
+- **Status, digest, end-of-day review.** A status block at session start and on each prompt, a daily archive, `hubkit digest` for when you come back, `hubkit eod`, a per-class `hubkit trust` ledger, and `hubkit inbound ingest` to track asks from other people so none goes unanswered.
 
 `SKILL.md` is the behaviour contract the hub session reads. This file is the operator view.
 
 ## Requirements
 
-- Linux. hubkit has only been run on Linux; macOS and Windows code exists but has never been run (see [Platform support](#platform-support)).
+- Linux. hubkit has only been run on Linux so far (see [Platform support](#platform-support)).
 - Python 3.14 or newer (the code uses `except A, B:` without parentheses, PEP 758). Standard library only.
 - [Claude Code](https://code.claude.com), which provides the hooks and the skills directory.
 - Optional: `gh` (GitHub CLI) on `PATH` for PR checks; without it items show "PR unknown". Optional: [`mep`](https://pypi.org/project/miseenplace-cli/) for a personal task list in the status block; without it the block shows `mep: UNVERIFIED (mep CLI not found ...)`, and `[mep] enabled = false` silences that.
@@ -50,7 +51,7 @@ Install hubkit with `uv tool install hubkit-cli`, run `hubkit skills --install D
 for your tool, then run the hubkit-setup skill and follow it. Ask me only for things you cannot find out yourself.
 ```
 
-hubkit runs on Claude Code hooks, so the skills directory is Claude Code's: `~/.claude/skills` (user level) or `.claude/skills` (project level). The PyPI name `hubkit-cli` is free as of 2026-11-06 and is not published yet; until it is, install straight from GitHub with `uv tool install git+https://github.com/ankitkpandey1/hubkit` (or from a clone with `uv tool install ./hubkit`).
+hubkit runs on Claude Code hooks, so the skills directory is Claude Code's: `~/.claude/skills` (user level) or `.claude/skills` (project level). `hubkit-cli` is on [PyPI](https://pypi.org/project/hubkit-cli/). To install the latest commit instead: `uv tool install git+https://github.com/ankitkpandey1/hubkit`.
 
 `hubkit skills --install DIR` copies three skills into `DIR`: `hub` (the rules and the scripts the hooks run), `hub-mep` (optional) and `hubkit-setup` (finds your facts, writes the config, registers the hooks, binds, proves it with `hubkit doctor`). Run it again after an upgrade to refresh the copies. `hubkit hooks --skills-dir DIR` prints the hooks block for `~/.claude/settings.json`.
 
@@ -126,12 +127,13 @@ Every key is documented in [docs/configuration.md](docs/configuration.md).
 
 | Behaviour | Strength |
 |---|---|
-| The Stop hook holds the turn while the hub owes something: no watch running, a stalled item, an unbacked verdict, an unsolved blocker, an item over its estimate, a "done" while your newest prompt is unread | Enforced (blocks the turn end) |
+| The Stop hook holds the turn while the hub owes something: no watch running, a stalled item, an unbacked verdict, an unsolved blocker, an item over its estimate, a "done" while your newest prompt is unread, a claim of passing tests with no command run, an approval while CI is not green | Enforced (blocks the turn end) |
 | An order closes only with the evidence its kind requires | Enforced (the close is refused) |
-| A learned class counts as fixed only if a rule for it exists | Enforced (the address is refused) |
-| A second job sent to a busy owner while an order is open; a brief that bans the action a blocker needs | Warning only |
-| `gh pr merge` with no live order naming that PR (the merge guard, `hubkit hook pretool`) | Warning only, never denies |
-| `hubkit doctor` warning when your `intents.toml` lacks keys the example has | Warning only |
+| A learned class counts as fixed only if a rule for it exists (and, with `--check`, its command passes) | Enforced (the address is refused) |
+| A merge order closes only with the merge commit sha | Enforced (the close is refused) |
+| A second job sent to a busy owner while an order is open; a brief that bans the action a blocker needs; a brief with no done-when, output or estimate line; more than 4 agents holding work | Warning only |
+| `gh pr merge` with no live order naming that PR, a post without your typed yes, a banned command, `TaskStop`, a repeated command, a whole-suite `pytest`, an unmarked decision card (`hubkit hook pretool`) | Warning only, never denies |
+| `hubkit doctor` warning when your `intents.toml` lacks keys the example has, a hook registered twice, or no scheduled `hubkit eod` | Warning only |
 
 Details and exact rules: [SKILL.md](SKILL.md) and [docs/architecture.md](docs/architecture.md#stop-gate-gatepy).
 
@@ -154,19 +156,11 @@ python3 tests/run_reverts.py        # each guard reverted in a copy must fail it
 
 ## Platform support
 
-All OS-specific code is in one module, `scripts/hubcore/compat.py` (standard library only): process start time and parent (`/proc` on Linux, `ps` on macOS, kernel32 through `ctypes` on Windows), file locks (`flock`, or `msvcrt` on Windows), the scheduler check (`crontab -l`, or `schtasks /query` on Windows) and UTF-8 stdio.
-
-| OS | Status |
-|---|---|
-| Linux | Full test suite run and passing here (ruff, mypy, pytest). |
-| macOS | The `ps` code path is exercised on Linux against the real `ps` (`tests/test_compat.py`), and mypy checks it with `--platform darwin`. Never run on a Mac. Not claimed to work until someone runs it there. |
-| Windows | The `ctypes` and `msvcrt` code is type-checked with `--platform win32` only. Never run on Windows. Tests that need a POSIX `sh` shim for `gh`, `crontab` or a notifier are skipped there. Not claimed to work until someone runs it there. There is no default desktop notification on Windows: set `tools.notify_cmd`. |
-
-There is no CI. Update this table only from a real run on that OS. On Windows the schedule is a Task Scheduler task whose name or command holds `hubkit` (for example `schtasks /create /sc minute /mo 5 /tn hubkit-tick /tr "python %USERPROFILE%\.claude\skills\hub\scripts\hubkit tick --if-bound"`), and hooks in `settings.json` call `python ...\hubkit hook <event>`.
+hubkit has only been run on Linux so far, and the full test suite, ruff and mypy pass there. The code has paths for macOS and Windows, all in `scripts/hubcore/compat.py` (process lookup, file locks, scheduler check, UTF-8 stdio), but they have never been run on those systems. Treat them as unsupported until someone does.
 
 ## Limits and known gaps
 
-- Python 3.14 only. macOS and Windows are written but unproven; see Platform support.
+- Python 3.14 only, and Linux only so far; see Platform support.
 - It targets one hub per machine. A second live hub is refused at bind.
 - Order creation takes a blocking `flock` (`scripts/hubcore/orders.py`); a process that holds the lock forever stalls order opens.
 - A hook payload with no `origin` field is treated as human input.

{hubkit_cli-0.1.0 → hubkit_cli-0.2.0}/SKILL.md

@@ -43,7 +43,7 @@ A human prompt containing an order word (words come from `intents.toml`; hotfix,
 - **(a2) Saying it is blocked is not solving it.** A final message that calls an open order blocked, down, failed or waiting on a host, tool or person is held until a block record with a fix exists (word list: `intents.toml` `[stop_gate]`). *Enforced in code.*
 - **(b) An urgent step never queues behind other work on a busy agent.** Send it to a free owner. The routed hook warns when a second job goes to an owner already holding a live job while an order is open. *Warning enforced in code; the choice of owner is prose.*
 - **(f) Re-read his words before saying done.** `hubkit asks ls --since 12h` prints his recent prompts one per line; a final message that says done while his newest prompt is newer than the last `work done` and the last `asks ls` is held. *Enforced in code.*
-- **(g) A merge is done on his order only.** `hubkit hook pretool` (register it as a PreToolUse hook on Bash) warns, never denies, on `gh pr merge` with no live order and none of his typed words naming that PR. *Warning only.*
+- **(g) A merge is done on his order only.** `hubkit hook pretool` (register it with `hubkit hooks --merge-guard`) warns, never denies, on `gh pr merge` with no live order and none of his typed words naming that PR; the same hook also warns on posts without his typed yes, banned commands, stopping an agent and an unmarked decision card. *Warning only.*
 - **(c) An agent's "that is a person's call" is checked, not relayed.** Read what the person already decided (the record, the order's words) before asking or telling anyone it needs them. *Checked in code:* a final message that relays "a person's call" with no Read/Grep/Glob/Bash call in that turn is held at the Stop gate (words: `intents.toml` `[stop_gate] unchecked_words`).
 - **(d) Before saying something is gone or absent, read the code that writes it.** A missing file, row or log line is proof of nothing until you know what produces it and when. *Checked in code:* a final message that says something is gone, lost or absent, with no lookup in that turn, is held the same way.
 - **(e) A brief must not forbid the one action that clears the blocker.** Before sending a brief, check its banned-command list against the step the order needs. *Checked in code:* the routed hook warns when a brief's "do not / never" line shares a key word with an open order's blocker or fix.
@@ -80,7 +80,7 @@ Silence on routed work past `stall_min` (reviews: `review_stall_min`, hard `OVER
 - At most 15 lines. Lead with what changed or what the user must decide. Tag claims verified / read / inferred. Full URLs. Times in the user's zone.
 - One line names what was left out and where it is. Input needed: the question tool after the report, recommended option first.
 - Hand over all PRs of a wave in one list, only when reviewed and ready.
-- Code or commands for the user go to a file, not a raw block (the overlay may name the clipboard step).
+- Code or commands for the user go to a file, not a raw block (the overlay may name the clipboard step). A one-line resume command is the exception: put it in the message as typed.
 
 ## 8. End of day
 

{hubkit_cli-0.1.0 → hubkit_cli-0.2.0}/docs/architecture.md

@@ -46,7 +46,7 @@ State lives in `$HUBKIT_STATE`, else `$XDG_STATE_HOME/hubkit` (default `~/.local
 | PostToolUse `AskUserQuestion` | `answer` | A card answer goes through the same intake and human-only filter as a typed prompt, and closes `ask` work items. |
 | PostToolUse `SendMessage\|Agent` | `routed` | Records or refreshes a routed-work item (see Routed work). One `gh` baseline read per named PR, 3 s in total. |
 | Stop | `stop` | The gate. Prints JSON `{"decision": "block", "reason": ...}` while the hub owes something. |
-| PreToolUse `Bash` | `pretool` | Warn-only merge guard: `gh pr merge` with no live order and none of his typed words naming the PR adds a warning as additional context. Never denies. Register it yourself in `settings.json`. |
+| PreToolUse `Bash\|TaskStop\|AskUserQuestion` | `pretool` | Warn-only guard; the warning goes in as additional context and nothing is ever denied. It warns on: `gh pr merge` with no live order and none of his typed words naming the PR; a post to a person or thread (`mep send|reply|post`, a chat send, `gh pr|issue comment|review`, a `gh api` comments or reviews call, or a Slack post or send) with no typed yes of his in the last 10 minutes; a banned command (`rm -rf`, `pkill`, `git reset --hard`, `git stash`, `git clean`, a forced push, `git checkout <ref> -- <path>`); the same command run 3 times in 10 minutes; `pytest` with no path, `-k` or `-m`; `TaskStop` without a stop word of his in the last 30 minutes; a decision card whose first option is not marked (Recommended). Register it with `hubkit hooks --merge-guard`, or add it to `settings.json` yourself. |
 | any | `record` | Schema canary. Every hub run writes the real payload's keys and value types (never values) to `payload-samples/`. |
 
 A hook failure is shown and logged (`hook-errors.jsonl`) only for the hub. The stop gate's own blocks are not counted as failures.
@@ -59,7 +59,7 @@ An order is a row family in `orders.jsonl` (`open`, `move`, `hold`, `resume`, `c
 - **Clock.** A first action is due within the kind's `first_action_min`; after that, movement (`order move`, or a touched `movement_files` glob) is due every `stall_min`. A deadline in the prompt ("in 30 min", "by 17:00") is parsed too.
 - **Holds.** Only explicit `hold_words` hold an order; `urgent_words` keep or restart the clock; `cancel_words` cancel. Long prompts (`verb_max_chars`) never hold or cancel.
 - **Chases.** A short prompt matching `intake.chase` while an order is open is recorded as a chase. Three chases print a stop-loss line.
-- **Close.** `order close` requires every evidence name the kind's `done_when` lists; `file:<path>` evidence must exist and be newer than the order. A closed order is archived under `archive/archive_<day>.jsonl`.
+- **Close.** `order close` requires every evidence name the kind's `done_when` lists; `file:<path>` evidence must exist and be newer than the order. Evidence names in the kind's `sha_evidence` (the merge kind's `merged`) must hold a commit sha. A closed order is archived under `archive/archive_<day>.jsonl`.
 
 ## Routed work (`routed.py`)
 
@@ -98,6 +98,9 @@ Every tick (cron with `--if-bound`, or inside `hubkit watch`) also checks things
 - **Notification.** New alerts go to `tools.notify_cmd` when set. With none set, only `hub:*` and `peer:*` alerts use the platform default (`notify-send` on Linux, `osascript` on macOS, when installed). Nothing here ever posts to chat.
 - **False alarms.** A silent item is not a stall while it is `blocked_on` another live job (`hubkit work block`), or while its holder's session is `busy` and its transcript is still growing (for up to 3x the limit). A head moved by the owner's own newest commit is progress, not a moved head. Every stall alert names the holder's state (`holder: ...`).
 - **Short asks.** A short message to a peer is tracked when a request verb opens a sentence or it ends in a question; acknowledgements, status questions and thanks are not.
+- **Own work.** `active` and `priority` items need an estimate within 5 minutes, say so out loud when over it, and raise a stale-verification alert when not stamped with `hubkit work verify` for `watcher.verify_stale_min`.
+- **Finished agents.** A background agent that reported finished more than `watcher.stale_agent_min` ago while its item is still listed raises `agent:<owner>:stale`.
+- **Inbound asks.** `hubkit inbound ingest` opens one `reply` item per message id from the inbox watcher's lines; one owed longer than `watcher.reply_owed_min` raises one alert.
 - **PRs are keyed by repo and number.** `owner/repo#N`, a full URL, or `<alias> N` for a `routed.repo_aliases` entry name their repo; only a bare `#N` falls back to `routed.repo`.
 
 ## Stop gate (`gate.py`)
@@ -110,7 +113,12 @@ The Stop hook blocks the end of the turn, with the exact next action, while any
 4. A job with an estimate on record is past it and not yet addressed (see the slowness check).
 5. A blocker is not being solved (no fix running, fix past its limit), a STOP-LOSS has no result since, or the final message calls an open order blocked/down/failed/waiting while no order has a fix on record.
 6. The final message says done while his newest prompt is newer than the last `work done` and the last `hubkit asks ls`; or it calls something gone/lost/absent, or "a person's call", with no Read/Grep/Glob/Bash call in the turn (transcript read).
-7. The final message ends on a question (or holds a `stop_guard.phrases` phrase) and no `ask` work item says anything really waits on the user.
+7. The final message ends on a question (or holds a `stop_guard.phrases` phrase) and no `ask` work item says anything really waits on the user. A question that is really blocked on him must use the card tool, not prose. An `ask` item asked 3 times with no answer also holds the turn: take the safest default and record it.
+8. The final message says "tests pass", "it works" or "is deployed" (`[stop_gate] verified_words`) but the turn ran no Bash command; names a PR or issue number with no link (`[stop_gate] full_links`); approves a PR while CI at its head is not green; or reports built work while the repo at the hook's `cwd` has uncommitted changes or unpushed commits.
+
+## Digest and trust
+
+`hubkit digest [--since 12h]` prints what finished, which orders closed, the latest alerts and what waits on the user. `hubkit trust record --class C --outcome accept|edit|reject` appends to `trust.jsonl`; `trust ls` shows each class as `auto` (enough accepts and nothing else) or `ask` (propose first), and the session status block names the classes that earned `auto` and those he rejected.
 
 ## The learn loop
 
@@ -119,7 +127,7 @@ This is the adaptive part. The mechanism, in the order it runs (`hooks.py`, `lea
 1. **Detect.** The prompt hook injects a `CORRECTION` line, with the quote, when a typed prompt matches a `[corrections] patterns` regex or repeats a recent prompt (word-set similarity at least `repeat_similarity` within `repeat_window_min`).
 2. **Record** (`learn.py add`). The hub records a class of mistake: kebab-case class, the gate it broke (G1 to G8 in SKILL.md), the exact quote, situation, wrong, right, evidence. A repeat of an existing class appends a dated `seen` line. Each entry stores a `basis` hash of the rule texts (overlay plus SKILL.md) at the time. Every field is redacted; edits are locked and atomic.
 3. **Promote** (`learn.py promote-candidates`, then `learn.py promote <class> --gate G --rule "..."`). A class seen two or more times is a candidate. Promoting writes one short rule line tagged `[<class>]` into your overlay under its gate. Personal rules stay out of the shared `SKILL.md`.
-4. **Address** (`learn.py address --class <c> --how "..."`). The class is marked fixed only if the rules changed since the entry was last seen **and** a rule line tagged `[<class>]` exists. Otherwise it is refused: a fix without a rule is not a fix.
+4. **Address** (`learn.py address --class <c> --how "..."`). The class is marked fixed only if the rules changed since the entry was last seen **and** a rule line tagged `[<class>]` exists. Otherwise it is refused: a fix without a rule is not a fix. With `--check CMD` the command must also exit 0, and every earlier check must still pass. `promote` is refused when the overlay would pass 200 lines. A class seen again after it was addressed is marked ineffective.
 5. **Probation.** The status block shows at most five learnings seen in the last 30 days; an addressed class stays visible for 14 days as "probation". If the class is seen again after it was addressed, `hubkit eod` and `learn.py review` list it as regressed (a fix that did not work), and it is to be fixed the same day.
 
 hubkit does the detecting, storing, refusing and reporting; the hub model does the recording and the rule writing, so the loop only adapts as far as the hub follows it.

{hubkit_cli-0.1.0 → hubkit_cli-0.2.0}/docs/commands.md

@@ -7,7 +7,7 @@
 | Verb | What it does |
 |---|---|
 | `skills --install DIR` | Copies the hub, hub-mep and hubkit-setup skills into DIR (re-run refreshes; the running copy is left alone). |
-| `hooks [--skills-dir DIR] [--merge-guard]` | Prints the `hooks` JSON for settings.json. `--merge-guard` adds the warn-only PreToolUse merge guard. Never edits settings.json. |
+| `hooks [--skills-dir DIR] [--merge-guard]` | Prints the `hooks` JSON for settings.json. `--merge-guard` adds the warn-only PreToolUse guard (matcher `Bash|TaskStop|AskUserQuestion`, see [architecture](architecture.md#hooks-scriptshubcorehookspy)). Never edits settings.json. |
 | `hubkit bind [--pid N]` | Bind this session (found by walking up the process tree, or the process `N`) as the hub. Interactive sessions only; refuses while another hub is live. |
 | `hubkit unbind` | Remove the bind. |
 | `hubkit status [--json]` | Rebuild `status.json` and print the status block (or the JSON). |
@@ -16,6 +16,10 @@
 | `hubkit order open --kind K --quote "..." [--unconfirmed]` | Open an order by hand. |
 | `hubkit order confirm ID` | Turn an unconfirmed order into an open one. |
 | `hubkit order move ID --note "..." [--estimate N]` | Record real movement. The first move also records the order's estimate (N minutes, else the usual time for its kind; printed when defaulted). |
+| `hubkit digest [--since 12h]` | What finished, which orders closed, the latest alerts and what waits on the user, for a user coming back. |
+| `hubkit inbound ingest [--file F]` | Turns the inbox watcher's lines into `reply` work items (one per asker, no duplicates). |
+| `hubkit trust record --class C --outcome O` / `trust ls` | Per-class record of approved and corrected outcomes; a class with repeated corrections is shown as ask-first. |
+| `hubkit work verify ID` | Stamp the item as checked against reality just now. An `active` or `priority` item not verified for `watcher.verify_stale_min` raises an alert. It runs no check itself. |
 | `hubkit asks ls [--since 12h]` | His recent prompts (from `recent-prompts.jsonl`), one line each, newest last. Marks the asks as re-read for the Stop gate. |
 | `hubkit order block ID --on "..." [--fix "..."] [--holder who]` | Record what blocks an order and the fix running now (a second call with the same `--on` adds the fix). No fix, or a fix older than `blocker_fix_min` with no result, alerts and holds the Stop gate. |
 | `hubkit order result ID --note "..."` | Record the result of an action: movement, clears the blocker and any STOP-LOSS. Notes from `order move` do neither. |
@@ -34,8 +38,8 @@
 | `hubkit tick [--if-bound] [--no-notify]` | One watchdog pass; prints `ALERT ...` lines. `--if-bound` is a no-op unless a hub is bound or an order is open (for cron). |
 | `hubkit watch [--interval S] [--max-cycles N]` | Loop `tick` (defaults 300 s, 12 cycles). Exits 0 printing `ALERT ...`, 3 after the cycles, 4 if another watch runs. |
 | `hubkit eod [--day D] [--note ".."]` | Write and print the end-of-day review. |
-| `hubkit doctor [--check-clean]` | Print config, state, bind, hook heartbeats, ledger health and a count of secret-shaped strings in state. `--check-clean` also lists checkout files containing a `private.terms` word. Exit 1 on a problem. |
-| `hubkit hook EVENT` | Run one hook (`session`, `prompt`, `answer`, `routed`, `stop`, `record`); reads the payload from stdin. For Claude Code, not for people. |
+| `hubkit doctor [--check-clean]` | Print config, state, bind, hook heartbeats, ledger health and a count of secret-shaped strings in state. Warns when your `intents.toml` lacks keys the example has, when a hubkit hook is registered twice in `settings.json` (read only), and when no scheduler entry runs `hubkit eod`. `--check-clean` also lists checkout files containing a `private.terms` word. Exit 1 on a problem. |
+| `hubkit hook EVENT` | Run one hook (`session`, `prompt`, `answer`, `routed`, `stop`, `record`, `pretool`); reads the payload from stdin. For Claude Code, not for people. |
 
 ## learn.py
 
@@ -47,8 +51,8 @@
 | `list [--last N]` | The N most recently seen entries (default 30). |
 | `promote-candidates [--min N]` | Classes seen at least N times (default 2). |
 | `promote CLASS --gate G --rule "..."` | Write a rule line tagged `[CLASS]` into the overlay under its gate. |
-| `address --class C --how "..."` | Mark a class fixed. Refused unless the overlay or SKILL.md changed since it was last seen and holds a `[C]` rule line. |
-| `review [--day D]` | New, repeated, regressed, holding and unaddressed classes for a day. |
+| `address --class C --how "..." [--check CMD]` | Mark a class fixed. Refused unless the overlay or SKILL.md changed since it was last seen and holds a `[C]` rule line. `--check` also runs a command that must exit 0 first, and every earlier `--check` must still pass. Promote is refused past the overlay's 200-line cap. |
+| `review [--day D]` | New, repeated, regressed, holding and unaddressed classes for a day, plus rules to expire (addressed and not seen since) and rules that exist only as prose. A class seen again after it was addressed is marked ineffective. |
 
 ## Example
 

{hubkit_cli-0.1.0 → hubkit_cli-0.2.0}/docs/configuration.md

@@ -81,6 +81,9 @@ Estimates and the slowness check (hub-internal, see architecture).
 | `waiting_min` | `3` | A peer waiting on a prompt this long raises `peer:<name>:waiting`. |
 | `active_s` | `120` | A busy holder whose transcript grew this recently is working, not stalled. |
 | `watch_grace_min` | `5` | Live work with no running `hubkit watch` this long raises `watch:dead`. |
+| `stale_agent_min` | `30` | A background agent that reported finished this long ago while its item is still listed raises `agent:<owner>:stale`. |
+| `reply_owed_min` | `60` | A `reply` item (from `hubkit inbound ingest`) open this long raises one alert. |
+| `verify_stale_min` | `180` | An `active` or `priority` item not stamped with `hubkit work verify` for this long raises an alert. |
 | `notify_user` | `true` | Use the platform default desktop notification for hub and peer alerts when `tools.notify_cmd` is empty. |
 
 ### `[stop_guard]`
@@ -101,6 +104,7 @@ Estimates and the slowness check (hub-internal, see architecture).
 | `review_stall_min` | `5` | Reviews: chase after this. |
 | `review_overdue_min` | `10` | Reviews: hard OVERDUE this long after routing, whatever the owner said. |
 | `skip_owners` | `["inbox"]` | Owners that are never tracked. |
+| `max_agents` | `4` | More distinct agents than this holding live routed work makes the routed hook warn. |
 | `min_chars` | `200` | A message without a PR is work only from this length. |
 | `gh_timeout_s` | `10` | Timeout for one `gh` read. |
 | `queued_wait_min` | `60` | Work queued "after X" gets no review clock until this long after the ask (or until a review shows). |
@@ -126,11 +130,20 @@ The first kind whose words match wins, so put the narrower kind first (`hotfix`
 | `not_order_phrases` | Phrases that contain an order word but are not orders ("release notes"). |
 | `question_start` | Regex; an order word in a question is never an order, urgent words or not (the hook then says so and shows `hubkit order open` for opening one by hand). |
 | `unconfirmed_ttl_min` | `tick` expires unconfirmed orders after this long. |
+| `blocker_fix_min` | A blocker with no fix, or a fix with no result after this many minutes (default `3`), alerts and holds the Stop gate. |
 | `held_review_min` | A hold this long raises an alert. |
 | `cancel_words`, `resume_words` | Words for the cancel and resume verbs. |
 | `verb_max_chars` | Prompts longer than this never hold, cancel, resume or count as chases. |
 | `chase` | Regex; a short prompt matching it while an order is open is a chase. |
 
+### `[stop_gate]`
+| Key | Meaning |
+|---|---|
+| `blocked_words` | A final message with one of these, while an order is open and none has a fix on record, cannot end the turn. |
+| `unchecked_words` | "A person's call" style claims held when the turn made no lookup. |
+| `verified_words` | "Tests pass", "it works", "is deployed" style claims held when the turn ran no Bash command. |
+| `full_links` | Default `true`. Set `false` to stop holding messages that name a PR or issue number with no link. |
+
 ### `[corrections]` (the learn loop)
 | Key | Meaning |
 |---|---|
@@ -146,6 +159,7 @@ The first kind whose words match wins, so put the narrower kind first (`hotfix`
 | `stall_min` | No movement for this long raises an alert. |
 | `dispatch` | Text printed when the order opens. |
 | `done_when` | Evidence names `order close` requires (`--evidence name=value`). |
+| `sha_evidence` | Evidence names that must hold a commit sha (the example merge kind uses `["merged"]`). |
 | `movement_files` | Globs; a file touched after the order opened counts as movement. |
 
 Add your own kinds the same way; the words are matched as whole words, case-insensitively.

{hubkit_cli-0.1.0 → hubkit_cli-0.2.0}/intents.example.toml

@@ -37,7 +37,7 @@ blocked_words = ["blocked", "is down", "are down", "failed", "failing", "no host
 [corrections]
 # A human prompt matching one of these regexes (or repeating the previous prompt) makes the prompt
 # hook inject a reminder to record it: learn.py add with the quote.
-patterns = ["(?i)\\b(useless|stupid|dumb|idiot|wtf|r u (dumb|kidding|sleeping))\\b", "(?i)^\\W*(wrong|no+|nope)[!. ]*$", "(?i)\\bwhy (did|didn't|do|are) (you|u)\\b", "(?i)\\bi (told|asked|said)\\b", "(?i)\\balready (told|said|asked|replied)\\b", "(?i)\\bnot what i (asked|said|meant)\\b"]
+patterns = ["(?i)\\b(you|u|ur|you're|youre|r u)\\s+(are\\s+|r\\s+)?(so\\s+|really\\s+)?(useless|stupid|dumb|idiot|kidding|sleeping)\\b", "(?i)^\\W*(useless|stupid|dumb|idiot|wtf)\\W*$", "(?i)\\bwtf\\b", "(?i)^\\W*(wrong|no+|nope)[!. ]*$", "(?i)\\bwhy (did|didn't|do|are) (you|u)\\b", "(?i)\\bi (told|asked|said)\\b", "(?i)\\balready (told|said|asked|replied)\\b", "(?i)\\bnot what i (asked|said|meant)\\b"]
 repeat_similarity = 0.6
 repeat_window_min = 30
 
@@ -63,6 +63,7 @@ first_action_min = 5
 stall_min = 30
 dispatch = "Merge exactly the named set."
 done_when = ["merged"]
+sha_evidence = ["merged"]   # evidence names that must hold the merge commit sha
 movement_files = []
 
 [kinds.code_red]

{hubkit_cli-0.1.0 → hubkit_cli-0.2.0}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "hubkit-cli"
-version = "0.1.0"
+version = "0.2.0"
 description = "Session supervisor for Claude Code: hooks-based multi-agent orchestration with order clocks, routed-work tracking, a watchdog and a stop gate, standard library only."
 readme = "README.md"
 requires-python = ">=3.14"

{hubkit_cli-0.1.0 → hubkit_cli-0.2.0}/scripts/hubcore/cli.py

@@ -4,12 +4,14 @@ from __future__ import annotations
 
 import argparse
 import json
+import os
+import re
 import sys
 import time
 import tomllib
 from pathlib import Path
 
-from . import compat, config, hooks, install, lessons, routed, status, store, watcher
+from . import compat, config, hooks, inbound, install, lessons, routed, status, store, trust, watcher
 from .orders import Orders
 from .redact import count_hits_text
 
@@ -81,6 +83,8 @@ def cmd_work(a, cfg: dict) -> int:
             from . import slow
 
             print(slow.re_estimate(cfg, a.id, a.minutes or 0, a.why))
+        elif a.action == "verify":
+            print(status.work_verify(cfg, a.id))
         elif a.action == "done":
             print(status.work_done(cfg, a.id, a.note))
         else:
@@ -122,6 +126,29 @@ def intents_gaps(cfg: dict) -> list[str]:
     return gaps
 
 
+def duplicate_hooks(settings: Path) -> list[str]:
+    """Events for which the user's settings.json runs `hubkit hook <event>` more than once (each run double-counts). Read only."""
+    if not settings.exists():
+        return []
+    try:
+        hooks_cfg = json.loads(settings.read_text(encoding="utf-8")).get("hooks", {})
+    except ValueError, AttributeError:
+        return []
+    seen: dict[str, int] = {}
+    for blocks in hooks_cfg.values():
+        for block in blocks if isinstance(blocks, list) else []:
+            for h in block.get("hooks", []):
+                m = re.search(r"hubkit\S*\s+hook\s+(\w+)", str(h.get("command", "")))
+                if m:
+                    seen[m.group(1)] = seen.get(m.group(1), 0) + 1
+    return sorted(ev for ev, n in seen.items() if n > 1)
+
+
+def eod_scheduled() -> bool:
+    """True when a scheduler entry runs `hubkit ... eod`."""
+    return any("hubkit" in x and re.search(r"\beod\b", x) for x in compat.scheduler_lines())
+
+
 def cmd_doctor(a, cfg: dict) -> int:
     """Print config, bind, hook heartbeat and state health; exit 1 on secret-shaped strings or private terms."""
     sd = store.state(cfg)
@@ -138,6 +165,11 @@ def cmd_doctor(a, cfg: dict) -> int:
     print(f"orders.jsonl: {len(rows)} rows, {bad} bad lines")
     errs = store.read_rows(sd / "hook-errors.jsonl")[0]
     print(f"hook-errors.jsonl: {len(errs)} rows")
+    if not eod_scheduled():
+        print(f"WARNING no {compat.scheduler_name()} entry runs `hubkit eod`: the end-of-day review and its learning pass only happen when someone asks for them.")
+    dup = duplicate_hooks(Path(os.environ.get("HUBKIT_SETTINGS") or Path.home() / ".claude" / "settings.json"))
+    if dup:
+        print(f"WARNING settings.json runs hubkit hook more than once for: {', '.join(dup)}. Each run double-counts prompts and answers; remove the duplicate entries.")
     gaps = intents_gaps(cfg)
     if gaps:
         print(f"WARNING intents.toml lacks what intents.example.toml has: {', '.join(gaps)}. Copy those lines from the example; the checks they configure use built-in defaults or stay off.")
@@ -183,6 +215,15 @@ def main(argv: list[str] | None = None) -> int:
     hk.add_argument("--skills-dir", default="$HOME/.claude/skills")
     hk.add_argument("--merge-guard", action="store_true", help="include the warn-only PreToolUse merge check")
 
+    tr = sub.add_parser("trust")
+    tr.add_argument("action", choices=["record", "ls"])
+    tr.add_argument("--class", dest="cls", default="")
+    tr.add_argument("--outcome", default="")
+
+    ib = sub.add_parser("inbound")
+    ib.add_argument("action", choices=["ingest"])
+    ib.add_argument("--file", help="the inbox watcher's output (default: stdin)")
+
     k = sub.add_parser("asks")
     k.add_argument("action", choices=["ls"])
     k.add_argument("--since", default="12h", help="how far back, for example 30m, 12h, 2d")
@@ -202,7 +243,7 @@ def main(argv: list[str] | None = None) -> int:
     o.add_argument("--estimate", type=int, help="move: minutes this order should take (default: the usual time)")
 
     w = sub.add_parser("work")
-    w.add_argument("action", choices=["add", "done", "ls", "estimate", "block"])
+    w.add_argument("action", choices=["add", "done", "ls", "estimate", "block", "verify"])
     w.add_argument("id", nargs="?")
     w.add_argument("--kind", default="active")
     w.add_argument("--title", default="")
@@ -224,6 +265,8 @@ def main(argv: list[str] | None = None) -> int:
 
     st = sub.add_parser("status")
     st.add_argument("--json", action="store_true")
+    dg = sub.add_parser("digest")
+    dg.add_argument("--since", default="12h", help="how far back, for example 30m, 12h, 2d")
     e = sub.add_parser("eod")
     e.add_argument("--day")
     e.add_argument("--note", default="")
@@ -254,6 +297,10 @@ def main(argv: list[str] | None = None) -> int:
             print("\n".join(install.install_skills(Path(a.install).expanduser())))
         elif a.cmd == "hooks":
             print(install.hooks_block(a.skills_dir, a.merge_guard))
+        elif a.cmd == "trust":
+            print(trust.record(cfg, a.cls, a.outcome) if a.action == "record" else "\n".join(f"{c}: {m}" for c, m in sorted(trust.mode(cfg).items())) or "no outcomes recorded")
+        elif a.cmd == "inbound":
+            print("\n".join(inbound.ingest(cfg, Path(a.file).read_text(encoding="utf-8") if a.file else sys.stdin.read())) or "nothing new")
         elif a.cmd == "asks":
             print("\n".join(lessons.asks_lines(cfg, a.since)))
         elif a.cmd == "watch":
@@ -261,6 +308,8 @@ def main(argv: list[str] | None = None) -> int:
         elif a.cmd == "status":
             st_ = status.write_status(cfg)
             print(json.dumps(st_, indent=1) if a.json else status.render_block(cfg, Orders(cfg), st_))
+        elif a.cmd == "digest":
+            print("\n".join(status.digest(cfg, a.since)))
         elif a.cmd == "eod":
             day = a.day or store.day_of(cfg, time.time())
             out = status.eod(cfg, day, a.note)

{hubkit_cli-0.1.0 → hubkit_cli-0.2.0}/scripts/hubcore/config.py

@@ -28,14 +28,14 @@ DEFAULTS: dict = {
         "extensions_dir": "",  # default: <config dir>/ext
     },
     "tools": {"mep": "mep", "notify_cmd": "", "gh": "gh"},
-    "routed": {"stall_min": 15, "min_chars": 200, "skip_owners": ["inbox"], "gh_timeout_s": 10,
+    "routed": {"stall_min": 15, "min_chars": 200, "max_agents": 4, "skip_owners": ["inbox"], "gh_timeout_s": 10,
                "repo": "", "reply_window_min": 60, "queued_wait_min": 60, "owner_logins": {}, "skip_briefs": ["mep watch", "inbox watch"],
         "activity_globs": [
             "~/.claude/projects/*/*/subagents/agent-{id}.jsonl",
             "/tmp/claude-*/*/*/tasks/{id}.output",
         ], "reviewer_logins": [], "review_words": ["review"], "review_stall_min": 5, "review_overdue_min": 10},
     "slow": {"min_samples": 3, "order_default_min": 30, "floor_min": 2, "tool_min": 2, "scan_kb": 2048},
-    "watcher": {"stale_agent_min": 30, "stuck_min": 30, "unhandled_min": 10, "waiting_min": 3, "active_s": 120, "watch_grace_min": 5, "notify_user": True},
+    "watcher": {"stale_agent_min": 30, "reply_owed_min": 60, "verify_stale_min": 180, "stuck_min": 30, "unhandled_min": 10, "waiting_min": 3, "active_s": 120, "watch_grace_min": 5, "notify_user": True},
     "mep": {"enabled": True, "timeout_s": 30},
     "watchdog": {"transcript_scan_hours": 24, "cron_stale_min": 15},
     "think": {"decisions": ""},  # path of the active slug's decisions.md; its last lines go in the status block

{hubkit_cli-0.1.0 → hubkit_cli-0.2.0}/scripts/hubcore/gate.py

@@ -65,6 +65,23 @@ def unverified_verdict(cfg: dict, last: str) -> list[str]:
     return sorted(routed.pr_num(u) for u in urls.values() if not _reviewed_at_head(cfg, u, timeout))
 
 
+def ci_not_green(cfg: dict, last: str) -> list[str]:
+    """PRs the message gives an approving verdict on while CI at the head is failing or pending. A PR is resolved from
+    its URL, a repo-qualified ref or `routed.repo`; a bare number is matched to a tracked item's PR."""
+    out: list[str] = []
+    timeout = float(routed.opts(cfg)["gh_timeout_s"])
+    tracked = [u for i in routed.items(cfg) for u in [*i["prs"], *i.get("queue", [])]]
+    for sentence in re.split(r"(?<=[.!?])\s+|\n", last):
+        if not (re.search(r"(?i)\bapprov", sentence) and (PR_REF.search(sentence) or REPO_PR.search(sentence)) and not AWAITING.search(sentence)):
+            continue
+        for ref in routed.refs(cfg, sentence):
+            for u in [ref] if ref.startswith("http") else [t for t in tracked if routed.same_pr(t, ref)]:
+                info = routed.pr_info(cfg, u, timeout)
+                if info and info.get("ci") in ("failing", "pending"):
+                    out.append(f"#{routed.pr_num(u)} ({info['ci']})")
+    return sorted(set(out))
+
+
 def _reviewed_at_head(cfg: dict, url: str, timeout: float) -> bool:
     """True when GitHub shows a non-author review at the head, or cannot be read (an unreadable PR proves nothing)."""
     info = routed.pr_info(cfg, url, timeout)
@@ -124,21 +141,32 @@ def blocker_reasons(cfg: dict, last: str, now: float) -> list[str]:
     return out
 
 
-def stop_reasons(cfg: dict, last: str, now: float | None = None, transcript: str | None = None) -> list[str]:
+def stop_reasons(cfg: dict, last: str, now: float | None = None, transcript: str | None = None, cwd: str | None = None) -> list[str]:
     """Every reason the turn cannot end yet, each with its exact action. Empty = the turn may end."""
     now = now or time.time()
+    ci_bad = ci_not_green(cfg, last)  # before prompt_lines: a reviewed item closes there
     reasons = routed.prompt_lines(cfg, now)
     reasons += blocker_reasons(cfg, last, now)
-    reasons += [r for r in (lessons.unswept_asks(cfg, last, now), lessons.unchecked_claim(cfg, last, transcript)) if r]
+    if r := lessons.unpushed_work(last, cwd):
+        reasons.append(r)
+    reasons += [r for r in (lessons.unswept_asks(cfg, last, now), lessons.unchecked_claim(cfg, last, transcript), lessons.unrun_claim(cfg, last, transcript)) if r]
+    if refs := lessons.bare_refs(cfg, last):
+        reasons.append(f"hubkit: your message names {', '.join(refs[:4])} with no link. NEXT: give the full URL of each PR or issue so he can click it")
     try:
         reasons += slow.gate_lines(cfg, now)
     except Exception as e:  # a failing estimate check must not wedge the hub's turn end
         store.append_row(store.state(cfg) / "hook-errors.jsonl", {"where": "slow.gate", "error": repr(e)[:200]})
     if open_work(cfg) and not watch_live(cfg):
         reasons.append(f"hubkit: nothing will wake you when routed work stalls. NEXT: start in the background: {WATCH_CMD}")
+    reasons += [f"hubkit: the ask {i['id']} ({i['title'][:50]}) was asked {i['asked']} times with no answer. Stop asking. NEXT: take the safest default, record it with the reason, "
+                f"and go on: hubkit work done {i['id']} --note \"defaulted to ...\"" for i in work_load(store.state(cfg)) if i.get("kind") == "ask" and int(i.get("asked", 1)) >= 3]
     for n in unverified_verdict(cfg, last):
         reasons.append(f"hubkit: your message gives a verdict on #{n}, but GitHub shows no review at its current head "
                        "(agent words are not evidence). NEXT: correct the message; relay a verdict only after the item closes itself")
+    if ci_bad:
+        reasons.append(f"hubkit: your message approves {', '.join(ci_bad)} but CI at the head is not green. NEXT: say CI is not green, or wait for it, before relaying an approval")
+    if (r := lessons.prose_ask(last, transcript, is_question(cfg, last) and blocked_on_user(cfg))):
+        reasons.append(r)
     if is_question(cfg, last) and not blocked_on_user(cfg):
         reasons.append("hubkit: you ended on a question to the user but nothing is blocked on him. NEXT: settle it from the record "
                        "or route it; only if it truly needs him run: hubkit work add --kind ask --title \"<the one question>\" --note \"<ordered work it blocks>\"")