Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
15 changes: 10 additions & 5 deletions docs/system-specs/modules/crew-mode.md
Original file line number Diff line number Diff line change
Expand Up @@ -3946,9 +3946,9 @@ every assembled prompt, offline; the fixtures and their goldens are in
A member-mode slot's transcript is the Bot's whole working record: the
`[auto-nudge cycle N]` turns of its patrol loop, `[Cron notification …]` and
`[Subagent completion event]` envelopes, every tool call, and the say-nothing
reply a quiet Heartbeat or `[Worker report]` turn ends on. A work-ledger patrol
wake with nothing for the user ends on one visible line of what the Bot did
(`patrol_reply`), which the chat pane shows. The Members page's chat pane shows **only what
reply a quiet Heartbeat, `[Worker report]` or work-ledger patrol turn ends on: a
patrol wake with no news for the user says nothing, and a finished task is
reported in full (the Bot's own patrol rule in `conductor.patrol`). The Members page's chat pane shows **only what
the Bot says to the user** and the user's own messages. Everything else stays in the slot's history and is filtered at render time by `filterCrewmateChat`
(`website/src/components/chat/crewmateBubbles.ts`), which `ChatPane` applies
when its host passes `crewmate`. Ordinary chats never pass it and are drawn
Expand Down Expand Up @@ -4746,11 +4746,16 @@ compaction, and the long-task and
team rules every turn needs: the item budget, one owner per file and one
integrator per output, merging two lines, reusing a finished worker only for the
next step of its own line, the plan reviewer, the `pr_merged` bar,
the hourly check on a quiet wake, routine incidents, reporting before a park, and
the one-line reply on a patrol wake that has nothing the person must see) are written once,
the hourly check on a quiet wake, routine incidents, and reporting before a park) are written once,
in `agent_parts/spec_prompts.py` `_LEDGER_CONDUCTOR_RULES`: the root charter is
assembled from them (`_CONDUCTOR_PROMPT_SECTIONS`) and `conductor.patrol` names
each as a placeholder (`context_assembly.member.ledger_rules_block`).
How each conductor replies to a patrol wake is NOT shared: conductor mode,
watched live, gets one line (`_CONDUCTOR_PATROL_REPLY`, also in the team lead's
`_SHARED_CONDUCTOR_RULES_BLOCK`), while a Bot, read once or twice a day, says
nothing unless there is news (its own rule in `conductor.patrol`). The shared
`wake_check` and goal-conductor step 6 point at "your patrol-reply rule", so each
side reads its own.
`test/test_mate_conductor_parity.py` roots at `crew_mode.crew_mode_agent()`, the
agent conductor mode runs, checks that every shared rule is in both installed
prompts, and compares the effective agents, not the spec
Expand Down
1 change: 1 addition & 0 deletions src/kiro_crew/agent.py
Original file line number Diff line number Diff line change
Expand Up @@ -180,6 +180,7 @@
session_agent_spec_path,
)
from kiro_crew.agent_parts.spec_prompts import (
_CONDUCTOR_PATROL_REPLY,
_CONDUCTOR_PROMPT_SECTIONS,
_CONDUCTOR_SYSTEM_PROMPT,
_CONDUCTOR_WORK_LEDGER_PROTOCOL,
Expand Down
25 changes: 16 additions & 9 deletions src/kiro_crew/agent_parts/spec_prompts.py
Original file line number Diff line number Diff line change
Expand Up @@ -313,7 +313,7 @@
re-read the board: for each open item, compare its elapsed time with its
`budget: <N>m` and look for the files its brief named. Re-plan now, and say why,
any item that is stalled, past budget with nothing to show, or on a wrong brief.
With nothing wrong, the one line a patrol wake gets says so.
With nothing wrong, reply as your patrol-reply rule says.

""",
"routine_incidents": """**Handle routine incidents from the goal, then report them.** A worker out of
Expand All @@ -330,7 +330,17 @@
nobody where the goal stands.

""",
"patrol_reply": """**A patrol wake gets one line.** On a turn your work-ledger patrol opened
"skill_first": """The `goal-conductor` skill carries the operating procedure — the work-item
tests, the dispatch steps, the patrol cycle, the stop conditions. Read it before
acting on a goal. When a turn carries a REINJECTED AFTER COMPACTION block, the
skill text you read is gone: read it again before you act. """,
}

#: Conductor mode's reply to a patrol wake. Not a shared rule: a person watches a
#: conductor-mode chat live, while a Bot's thread is read once or twice a day, so
#: a Bot's ``conductor.patrol`` carries its own rule (a wake with no news says
#: nothing). The team lead runs conductor mode, so it carries this one.
_CONDUCTOR_PATROL_REPLY = """**A patrol wake gets one line.** On a turn your work-ledger patrol opened
(a `[work-ledger wake]` or an `[auto-nudge cycle N]`, its forced hourly wake
included) that has nothing the person must see, reply with ONE short line saying
what you did, never `nothing_to_do`: "Extended the docs task's budget to 300m."
Expand All @@ -341,12 +351,7 @@
is reported in full. A Heartbeat or `[Worker report]` turn keeps its own quiet
end.

""",
"skill_first": """The `goal-conductor` skill carries the operating procedure — the work-item
tests, the dispatch steps, the patrol cycle, the stop conditions. Read it before
acting on a goal. When a turn carries a REINJECTED AFTER COMPACTION block, the
skill text you read is gone: read it again before you act. """,
}
"""


#: ``_CONDUCTOR_SYSTEM_PROMPT`` in reading order, one ``(name, text)`` per section.
Expand Down Expand Up @@ -546,7 +551,7 @@
("wake_check", _LEDGER_CONDUCTOR_RULES["wake_check"]),
("routine_incidents", _LEDGER_CONDUCTOR_RULES["routine_incidents"]),
("report_before_park", _LEDGER_CONDUCTOR_RULES["report_before_park"]),
("patrol_reply", _LEDGER_CONDUCTOR_RULES["patrol_reply"]),
("patrol_reply", _CONDUCTOR_PATROL_REPLY),
(
"tracker",
"""**Add a tracker when one patrol read no longer fits.** Your `compact` ledger
Expand Down Expand Up @@ -728,6 +733,8 @@
"The dispatch steps are 1. `work_ledger_record` `action=create`, 2. "
"`session_create` with `agent` set explicitly, 3. `bind`, 4. the seed.\n\n"
+ "\n\n".join(rule.strip() for rule in _LEDGER_CONDUCTOR_RULES.values())
+ "\n\n"
+ _CONDUCTOR_PATROL_REPLY.strip()
+ "\n"
)

Expand Down
4 changes: 2 additions & 2 deletions src/kiro_crew/builtin_skills/goal-conductor/SKILL.md
Original file line number Diff line number Diff line change
Expand Up @@ -575,8 +575,8 @@ Each cycle:
**When the lines show the item itself is wrong** (already done on main,
mis-scoped, needing a split, a bar that cannot be met), re-plan that item now
(Close the round).
6. **A patrol wake gets one line**, as your prompt's patrol-reply rule says,
and more only in the cases it names. Never stop bare after your tool
6. **Reply to a patrol wake as your prompt's patrol-reply rule says**, and
write more only in the cases it names. Never stop bare after your tool
calls: that keeps the empty-response recovery.

### Close the round
Expand Down
17 changes: 12 additions & 5 deletions src/kiro_crew/crewmate_prompts.json
Original file line number Diff line number Diff line change
Expand Up @@ -194,7 +194,7 @@
" People read this thread once or twice a day, so report in it",
" rarely and in one message: an outcome and its evidence (\"re: <the",
" thing>\"), several finished things together, never a stream of small",
" updates (a work-ledger patrol wake gets its one line, as item 5 says).",
" updates.",
"3. When stuck, climb this ladder in order, and genuinely try each rung:",
" (a) try a genuinely DIFFERENT approach — another tool, entry point, or",
" strategy, not the same command again;",
Expand All @@ -209,9 +209,7 @@
" waiting costs. Keep it short. Before sending, reread the draft as a",
" stranger with no context would, and rewrite until they could act on it.",
"5. A quiet cycle is a successful cycle. Report real signals — results, walls,",
" threshold crossings — never \"nothing new\". When you run a work-ledger",
" patrol, its wakes are the one exception: each gets the one line your",
" patrol rule asks for.",
" threshold crossings — never \"nothing new\".",
" Every word you write as text is shown to the user: before a tool",
" call, between tool calls, and after them. Text is never a private",
" scratchpad. So when you have nothing to say to the user, write",
Expand Down Expand Up @@ -274,7 +272,16 @@
"{wake_check}",
"{routine_incidents}",
"{report_before_park}",
"{patrol_reply}",
"**On a patrol wake, speak only for news.** A turn your work-ledger patrol",
"opened (a `[work-ledger wake]` or an `[auto-nudge cycle N]`, its forced hourly",
"wake included) speaks only when there is news for the person: a task finished,",
"got stuck or was dropped, a pull request merged, the person must act or decide,",
"or you decided something on their behalf (a trust boundary, scope, product",
"shape). A finished task is always reported in full, in that same turn, before",
"you close its item or call `autonudge_stop`. Anything else is not news: a wait",
"you already reported, \"still running\", \"on track\", \"checked, nothing changed\".",
"Then write no text at all and end with `nothing_to_do`.",
"",
"{close_item}",
"",
""
Expand Down
4 changes: 2 additions & 2 deletions test/test_context_composition_contract.py
Original file line number Diff line number Diff line change
Expand Up @@ -108,8 +108,8 @@ def _transcript(turns: int, *, size: int) -> list[dict]:
"critical_rules_channel": "a55668069ebcf5a45055d59028c2b57f39e86546b2639fe88385f765e7ef6257",
"critical_rules_subagent": "89db6c776715d0b6358e140b8b13afba64bfe287e6e862e57b978baa09844d1b",
"docs_section": "971b6008afbd9ff048c498c06cffecc6dafc7a58b0df75ff9db0b15cf9e26c09",
"member_how_you_work": "55f0c5287f910b65f2709ab7cd736ceb84036d73e814b31d5a1b976edc12ce5f",
"member_how_you_work_unavailable": "64639dfa262f98a68a7aa60140384ef5476edf25ea62eb1b2b9364618a43c2a8",
"member_how_you_work": "1e03ce13c439da54cd97e4555ad9353cc19fb1208a8f2e0e64cdf856fe23ef3b",
"member_how_you_work_unavailable": "a16008defce445de4721140a9b305a7c97f03270e7c797f4c04ae4bd7ff7644d",
"reply_answer_only": "f0bd694165ff81cdbe123048042070588b846bd43a224b31adc79d9f8768bfce",
"reply_concise": "5d3e25d89e682469332237e46373f0e9aac8516eefd63047134d1ec0575aac2a",
"response_preferences_frame": "08ee149ae459c828213d47643f1cdbd6661d8a77a2223b8be5d0aa6fbb6cdc18",
Expand Down
43 changes: 24 additions & 19 deletions test/test_mate_conductor_parity.py
Original file line number Diff line number Diff line change
Expand Up @@ -80,6 +80,7 @@
"prompt:sealed_crew_paths": "a crewmate's conductor runs on the default contract, config/prompt.md, which states the same sealed-path rule",
"prompt:tool_inventory": "lists the root conductor's narrowed tool set; a crewmate has the default tools and its member pieces name the crewmate-only ones",
"prompt:goal_changes": "round-boundary goal changes; a crewmate's goal is set through set_goal (member pieces)",
"prompt:patrol_reply": "a person watches a conductor-mode chat live, so a patrol wake there gets one line; a Bot's thread is read once or twice a day, so conductor.patrol carries its own rule: a wake with no news says nothing",
"prompt:tracker": "a tracker is for a fleet one compact read cannot hold; a crewmate's ledger is one level on one machine and its [Worker report] turns carry each worker's state",
}

Expand Down Expand Up @@ -499,27 +500,32 @@ def test_the_long_task_and_team_rules_are_shared() -> None:
assert set(_NEW_SHARED) <= crewmate_prompts.placeholders(_MATE_PIECE)


def test_a_patrol_wake_gets_one_line_on_both_conductors(specs) -> None:
"""The one-line patrol reply is a shared rule both installed conductors carry, and
no quiet-end wording tells a patrol wake to stay silent, so neither side is told
two things."""
def test_a_patrol_wake_gets_one_line_in_conductor_mode_and_none_on_a_bot(specs) -> None:
"""Each conductor replies to a patrol wake its own way: conductor mode, watched
live, gets one line; a Bot, whose thread is read once or twice a day, says
nothing unless there is news. Neither carries the other's rule."""
root_spec, _mate = specs
rule = agent._LEDGER_CONDUCTOR_RULES["patrol_reply"].strip()
assert "patrol_reply" in crewmate_prompts.placeholders(_MATE_PIECE)
for side, text in (
(_ROOT, root_spec["prompt"]),
("kirocrew-mate-conductor", _mate_ledger_text()),
):
assert rule in text, f"{side} lacks the patrol_reply rule"
assert "nothing_to_do" not in agent._LEDGER_CONDUCTOR_RULES["wake_check"]
assert "its wakes are the one exception" in _words(
root_rule = agent._CONDUCTOR_PATROL_REPLY.strip()
assert "patrol_reply" not in agent._LEDGER_CONDUCTOR_RULES
assert "patrol_reply" not in crewmate_prompts.placeholders(_MATE_PIECE)
assert root_rule in root_spec["prompt"]
assert root_rule in agent._SHARED_CONDUCTOR_RULES_BLOCK
mate_text = _mate_ledger_text()
assert root_rule not in mate_text
mate_words = _words(mate_text)
assert "on a patrol wake, speak only for news." in mate_words
assert "a finished task is always reported in full" in mate_words
assert "then write no text at all and end with nothing_to_do." in mate_words
assert "its wakes are the one exception" not in _words(
crewmate_prompts.text("conductor.how_you_work")
)
assert 'never post "nothing changed"' not in _skill_text("goal-conductor")
# The one line is for a wake with no news only: milestones keep their full
# report (and the root's in-chat board), and the Bot's Heartbeat and
# [Worker report] turns keep their own quiet end.
words = _words(rule)
# Both rule sets point at "your patrol-reply rule", so one shared wake check
# and one skill step serve both.
assert "reply as your patrol-reply rule says" in _words(
agent._LEDGER_CONDUCTOR_RULES["wake_check"]
)
assert "as your prompt's patrol-reply rule says" in _skill_text("goal-conductor")
words = _words(root_rule)
assert "a task starting, finishing or getting stuck is reported in full" in words
assert "a heartbeat or [worker report] turn keeps its own quiet end" in words
talking = dict(agent._CONDUCTOR_PROMPT_SECTIONS)["talking_to_the_person"]
Expand Down Expand Up @@ -659,7 +665,6 @@ def test_the_team_lead_check_sees_a_lost_rule(
"bar_pr_merged": "as your prompt's rule on landed code says",
"plan_reviewer": "as your prompt's plan-reviewer rule says",
"wake_check": "a wake with no news is a check, as your prompt's rule says",
"patrol_reply": "as your prompt's patrol-reply rule says",
"reuse_worker": "your prompt's worker-reuse rule",
"read_message_detail": "as your prompt's read-message rule says",
"patrol_final_report": "end the patrol as your prompt's patrol rules say",
Expand Down
Loading