This repository has been archived on 2026-08-04. You can view files and clone it, but cannot push or open issues or pull requests.
orchestrator/agent-team/run-team.py
Adam Moussa 3c8317b06e feat(agent-team): durable GitHub-issue intake de-dup (schema v2)
The intake poller de-duped ingested issues in an in-memory set that does
not survive a process restart. A scheduled/cron intake (each run a fresh
process) would therefore re-ingest every still-open labeled issue on every
run and spawn duplicate pipeline tasks. Since the box is read-only (no write
token to remove the intake label), durable de-dup is the only correct guard.

- schema v2: new ingested_issues(source, issue_id, ingested_at) table +
  issue_already_ingested / record_issue_ingested helpers; migrate() adds the
  table to a legacy v1 DB and restamps; init_db creates it.
- github_intake: pluggable IngestStore seam (in-memory default preserved for
  tests/one-off; durable build_ledger_ingest_store for production). Record is
  after start_task succeeds, so a failed intake stays retryable.
- run-team.py intake-github wires the ledger store keyed by github:owner/repo,
  making a scheduled timer idempotent across runs.

978 tests pass (ruff clean).
2026-06-22 16:58:45 -04:00

1231 lines
46 KiB
Python

#!/usr/bin/env python3
"""``run-team.py`` — R720 agent-team operator entry CLI (design §3.3.1, §7.1 P1).
This is the **entry CLI** named in the design (§2 "entry CLI ``run-team.py``";
§9 "operator CLI"). It is the small manual path over the durable
``pending_questions`` ledger that §3.3.1 ("Manual path") requires::
A small CLI over the ledger lets an operator list ``open``/``parked``
questions, re-deliver, force-expire, or answer on a task's behalf; a stuck
task parks rather than spins. Destructive CLI actions (force-expire,
answer-on-behalf, force-resume) are audit-logged and require an explicit
confirmation flag.
It imports the committed FOUNDATION contracts verbatim — it does not redefine
them:
* :mod:`agent_team.db.schema` — :func:`connect`, :func:`init_db`,
:func:`answer_question`, :func:`expire_question`, :func:`supersede_question`,
:data:`QUESTION_STATES`.
* :mod:`agent_team.state_store` — :func:`atomic_write` for the append-only,
crash-safe audit log of destructive actions (§6.7 discipline).
Per the build constraints this is **pre-deployment scaffolding**: it provisions
nothing, enables no live CI, and performs no network or rsync. It only reads and
mutates the local SQLite ledger and writes a local audit log.
Subcommands (P1 surface):
* ``init-db`` — create/upgrade the agent-team tables in the ledger DB
(idempotent; wraps :func:`init_db`).
* ``list`` — list ``open`` (default) or any-status pending questions; with
``--parked`` it lists questions whose status is read as parked context. Pure
read; no confirmation needed.
* ``show`` — print one question row by ``question_id``. Pure read.
* ``expire`` — force-expire an ``open`` question (DESTRUCTIVE: requires
``--confirm``; audit-logged). Maps to :func:`expire_question`.
* ``answer`` — answer a question on a task's behalf (DESTRUCTIVE: requires
``--confirm``; audit-logged). Maps to :func:`answer_question`.
* ``supersede`` — mark a stale question ``superseded`` (DESTRUCTIVE: requires
``--confirm``; audit-logged). Maps to :func:`supersede_question`.
* ``intake-github`` — poll a repo for labeled open issues and start one pipeline
task per not-yet-ingested issue (one pass). Reads issues + calls the committed
coordinator intake entry only; no CI, OIDC, or git/patch apply. Opt-in/inert.
* ``intake-checker`` — the Plane-1 -> Plane-2 cross-plane loop (P5): read one or
more checker report JSON files / a dir, select CONFIRMED findings at/above a
severity threshold (default ``high``), de-dup, and start one remediation task
per unique finding via the committed coordinator intake entry. Reads local
report JSON only; no CI, OIDC, network, or git/patch apply. Opt-in/inert.
* ``fix`` — plan a Plane-1 Tier-3 dep-bump fix for one confirmed dependency-cve
finding (spec via Claude + minimal bump patch via DeepSeek) and, in
``--dry-run``, print the spec + patch + the org-CI ``workflow_dispatch`` inputs
WITHOUT dispatching. Opt-in/inert: it binds no dispatcher and holds no write
token; live dispatch is provisioning-gated (the P3-live apply/verify surface).
Exit codes: ``0`` success, ``1`` operational failure (e.g. row not found, the
compare-and-set lost the race), ``2`` usage error (argparse).
"""
from __future__ import annotations
import argparse
import getpass
import json
import os
import sqlite3
import sys
from datetime import datetime, timezone
from pathlib import Path
from typing import Any, Sequence
# ``run-team.py`` lives in ``agent-team/`` next to the importable ``agent_team``
# package. The hyphenated filename cannot itself be imported, so when run as a
# script we make the sibling package importable without an editable install
# (mirrors tests/conftest.py).
_CLI_DIR = Path(__file__).resolve().parent
if str(_CLI_DIR) not in sys.path:
sys.path.insert(0, str(_CLI_DIR))
from agent_team.db.schema import ( # noqa: E402 (path bootstrap must precede)
QUESTION_STATES,
answer_question,
connect,
expire_question,
init_db,
reopen_question,
supersede_question,
)
from agent_team.transport.base import Transport # noqa: E402 (path bootstrap)
# Transport choices the start/serve/intake commands accept (§3.3.1 D10). All
# three now have a live human-gate adapter wired in _build_transport: ``slack``
# (slack_sdk), ``github`` (requests issue/PR comment), and ``claude_code`` (local
# file drop). Each builder defers its optional SDK / token resolution to call
# time, so a missing dep/credential fails loudly only when that transport is
# actually selected.
_TRANSPORT_CHOICES: tuple[str, ...] = ("slack", "github", "claude_code")
__all__ = [
"build_parser",
"main",
]
# Default ledger DB location. Kept out of the repo (the package .gitignore
# excludes ``state/`` and ``*.sqlite``) so durable state is never committed.
_DEFAULT_DB = _CLI_DIR / "state" / "agent_team.sqlite"
# Default audit log for destructive actions, alongside the ledger DB.
_DEFAULT_AUDIT_LOG = _CLI_DIR / "state" / "audit.log.jsonl"
# Columns selected for list/show rendering, in display order.
_QUESTION_COLUMNS: tuple[str, ...] = (
"question_id",
"thread_id",
"turn",
"status",
"transport",
"channel_ref",
"posted_at",
"deadline_at",
"answered_at",
"answered_via",
)
# Destructive subcommands that require ``--confirm`` and are audit-logged.
# ``force-resume`` is the design-named operator verb (§3.3.1/§6.6); ``supersede``
# is kept as its lower-level alias. ``redeliver`` is NOT here — it is idempotent
# and non-destructive (it only clears a delivery ref), though it is still
# audit-logged for provenance.
_DESTRUCTIVE_ACTIONS: frozenset[str] = frozenset(
{"expire", "answer", "supersede", "force-resume"}
)
def _utc_now_iso() -> str:
"""Return the current UTC time as an ISO-8601 string (audit timestamps)."""
return datetime.now(timezone.utc).isoformat()
def _default_operator() -> str:
"""Best-effort OS login for audit attribution (never an empty string).
A previous empty default left destructive actions non-attributable (the
audit record named no one). Defaulting to the OS login keeps the §3.3.1
"audit-logged AND attributable" guarantee even when --operator is omitted;
falls back to "unknown" only if the login cannot be resolved.
"""
try:
user = getpass.getuser()
except Exception: # noqa: BLE001 - getuser can raise on odd environments
return "unknown"
return user or "unknown"
def _row_to_dict(row: sqlite3.Row) -> dict[str, Any]:
"""Project a ``pending_questions`` row to a plain dict for display."""
return {col: row[col] for col in _QUESTION_COLUMNS if col in row.keys()}
def _append_audit(audit_log: Path, entry: dict[str, Any]) -> None:
"""Append one JSON audit record with an atomic ``O_APPEND`` single write.
Destructive actions (§3.3.1) must leave an attributable trail. A previous
read-modify-rewrite design lost records under concurrent operators (two
processes each read the same bytes and the last rewrite wins). Instead each
record is one line written with ``O_APPEND``: the kernel serializes the
append and a write below ``PIPE_BUF`` is atomic on POSIX, so concurrent
appends never clobber each other. The file is created mode ``0600``
(operator identity / action content is sensitive) and re-chmod'd in case it
pre-existed wider. A failure here raises ``OSError`` BEFORE any ledger
mutation, preserving the audit-before-mutate guarantee.
"""
audit_log = Path(audit_log)
audit_log.parent.mkdir(parents=True, exist_ok=True)
line = (json.dumps(entry, sort_keys=True) + "\n").encode("utf-8")
fd = os.open(str(audit_log), os.O_WRONLY | os.O_CREAT | os.O_APPEND, 0o600)
try:
os.write(fd, line)
finally:
os.close(fd)
os.chmod(audit_log, 0o600)
def _audit_attempt(
audit_log: Path,
action: str,
*,
question_id: str,
operator: str,
detail: dict[str, Any] | None = None,
) -> None:
"""Record the *intent* to perform a destructive action BEFORE it mutates.
§3.3.1 requires every destructive action to be audit-logged. Writing the
attempt before the ledger mutation closes the "mutation applied with no
audit record" gap: if this append fails (e.g. an unwritable audit path) it
raises before any ledger row is touched, so the action aborts cleanly with
nothing changed. The matching :func:`_audit_outcome` records what happened.
"""
_append_audit(
audit_log,
{
"ts": _utc_now_iso(),
"action": action,
"phase": "attempt",
"question_id": question_id,
"operator": operator,
**(detail or {}),
},
)
def _audit_outcome(
audit_log: Path,
action: str,
*,
question_id: str,
operator: str,
applied: bool,
detail: dict[str, Any] | None = None,
) -> None:
"""Record the *result* of a destructive action AFTER it ran.
Carries ``applied`` (did the compare-and-set change a row). Pairs with the
:func:`_audit_attempt` record written before the mutation, so even if this
outcome append fails the attempt already proves the action was made.
"""
_append_audit(
audit_log,
{
"ts": _utc_now_iso(),
"action": action,
"phase": "outcome",
"question_id": question_id,
"operator": operator,
"applied": applied,
**(detail or {}),
},
)
def _require_confirm(action: str, *, confirm: bool) -> None:
"""Raise unless a destructive ``action`` was explicitly confirmed.
Mirrors the §3.3.1 rule: force-expire, answer-on-behalf, and force-resume
are audit-logged AND require an explicit confirmation flag. Failing closed
here means a typo can never silently mutate a live task's ledger row.
"""
if action in _DESTRUCTIVE_ACTIONS and not confirm:
raise PermissionError(
f"refusing destructive action '{action}' without --confirm "
f"(force-expire / answer-on-behalf / supersede are gated, §3.3.1)"
)
def _fetch_question(conn: sqlite3.Connection, question_id: str) -> sqlite3.Row | None:
"""Return the ledger row for ``question_id`` or ``None`` if absent."""
return conn.execute(
"SELECT * FROM pending_questions WHERE question_id = ?",
(question_id,),
).fetchone()
# --------------------------------------------------------------------------- #
# Subcommand handlers. Each returns a process exit code (0 ok, 1 op failure).
# --------------------------------------------------------------------------- #
def _cmd_init_db(args: argparse.Namespace, *, out: Any) -> int:
"""Create/upgrade the agent-team tables (idempotent)."""
init_db(args.db)
print(f"initialized ledger DB at {args.db}", file=out)
return 0
def _cmd_list(args: argparse.Namespace, *, out: Any) -> int:
"""List pending questions, optionally filtered by status.
Default lists ``open`` questions (the operator's "what is waiting" view).
``--status STATE`` narrows to one lifecycle state; ``--all`` lists every
state. ``--parked`` is a convenience alias that surfaces the parked-task
context an operator chases: questions that are no longer ``open`` (answered
but never resumed, expired, or superseded) and so may back a parked task.
"""
conn = connect(args.db)
try:
if args.all:
rows = conn.execute(
"SELECT * FROM pending_questions ORDER BY thread_id, turn"
).fetchall()
elif args.parked:
placeholders = ",".join("?" for _ in _PARKED_STATES)
rows = conn.execute(
f"SELECT * FROM pending_questions WHERE status IN ({placeholders}) "
"ORDER BY thread_id, turn",
tuple(_PARKED_STATES),
).fetchall()
else:
rows = conn.execute(
"SELECT * FROM pending_questions WHERE status = ? "
"ORDER BY thread_id, turn",
(args.status,),
).fetchall()
finally:
conn.close()
payload = [_row_to_dict(row) for row in rows]
print(json.dumps(payload, indent=2, sort_keys=True), file=out)
return 0
def _cmd_show(args: argparse.Namespace, *, out: Any) -> int:
"""Print one question row by ``question_id`` (pure read)."""
conn = connect(args.db)
try:
row = _fetch_question(conn, args.question_id)
finally:
conn.close()
if row is None:
print(f"no such question: {args.question_id}", file=sys.stderr)
return 1
print(json.dumps(_row_to_dict(row), indent=2, sort_keys=True), file=out)
return 0
def _cmd_expire(args: argparse.Namespace, *, out: Any) -> int:
"""Force-expire an ``open`` question (destructive; audit-logged).
Audits the attempt BEFORE mutating so a mutation can never land without a
trail (§3.3.1); records the outcome after.
"""
_require_confirm("expire", confirm=args.confirm)
_audit_attempt(
args.audit_log, "expire", question_id=args.question_id, operator=args.operator
)
conn = connect(args.db)
try:
changed = expire_question(conn, question_id=args.question_id)
finally:
conn.close()
_audit_outcome(
args.audit_log,
"expire",
question_id=args.question_id,
operator=args.operator,
applied=changed,
)
if not changed:
print(
f"expire no-op: question {args.question_id} was not 'open' "
"(already answered/expired/superseded or absent)",
file=sys.stderr,
)
return 1
print(f"expired question {args.question_id}", file=out)
return 0
def _cmd_redeliver(args: argparse.Namespace, *, out: Any) -> int:
"""Clear an ``open`` question's ``channel_ref`` so it is re-posted (§3.3.1).
The design's "re-deliver" operator action. Re-delivery itself is performed
by the transport reconcile loop; clearing ``channel_ref`` makes that loop
re-post and record a fresh ref. Idempotent and non-destructive (the question
stays ``open``), so it needs no ``--confirm`` — but it is audit-logged for
provenance. Returns ``1`` if the question is absent or not ``open``.
"""
_audit_attempt(
args.audit_log,
"redeliver",
question_id=args.question_id,
operator=args.operator,
)
conn = connect(args.db)
try:
row = _fetch_question(conn, args.question_id)
if row is None:
applied = False
prior_ref = None
status = None
elif row["status"] != "open":
applied = False
prior_ref = row["channel_ref"]
status = row["status"]
else:
prior_ref = row["channel_ref"]
status = "open"
conn.execute(
"UPDATE pending_questions SET channel_ref=NULL WHERE question_id=?",
(args.question_id,),
)
applied = True
finally:
conn.close()
_audit_outcome(
args.audit_log,
"redeliver",
question_id=args.question_id,
operator=args.operator,
applied=applied,
detail={"prior_channel_ref": prior_ref},
)
if not applied:
reason = "absent" if status is None else f"status={status}, not open"
print(
f"redeliver no-op: question {args.question_id} ({reason}); "
"nothing to re-post",
file=sys.stderr,
)
return 1
print(
f"cleared channel_ref for {args.question_id}; reconcile loop will re-post",
file=out,
)
return 0
def _cmd_answer(args: argparse.Namespace, *, out: Any) -> int:
"""Answer a question on a task's behalf (destructive; audit-logged).
Uses the foundation first-answer-wins compare-and-set: succeeds only if the
question is still ``open``. ``--answer`` is stored verbatim as the answer
payload string; ``--via`` records the answering identity for the audit
trail. The audit log records the operator regardless of outcome.
"""
_require_confirm("answer", confirm=args.confirm)
via = args.via or f"cli:{args.operator}"
_audit_attempt(
args.audit_log,
"answer",
question_id=args.question_id,
operator=args.operator,
detail={"answered_via": via},
)
conn = connect(args.db)
try:
changed = answer_question(
conn,
question_id=args.question_id,
answer_json=args.answer,
answered_via=via,
)
finally:
conn.close()
_audit_outcome(
args.audit_log,
"answer",
question_id=args.question_id,
operator=args.operator,
applied=changed,
detail={"answered_via": via},
)
if not changed:
print(
f"answer no-op: question {args.question_id} was not 'open' "
"(already answered/expired/superseded or absent)",
file=sys.stderr,
)
return 1
print(f"answered question {args.question_id} (via {via})", file=out)
return 0
def _cmd_supersede(args: argparse.Namespace, *, out: Any) -> int:
"""Mark a stale ``open``/``answered`` question ``superseded`` (destructive)."""
_require_confirm("supersede", confirm=args.confirm)
_audit_attempt(
args.audit_log,
"supersede",
question_id=args.question_id,
operator=args.operator,
)
conn = connect(args.db)
try:
changed = supersede_question(conn, question_id=args.question_id)
finally:
conn.close()
_audit_outcome(
args.audit_log,
"supersede",
question_id=args.question_id,
operator=args.operator,
applied=changed,
)
if not changed:
print(
f"supersede no-op: question {args.question_id} was not "
"'open'/'answered' (already expired/superseded or absent)",
file=sys.stderr,
)
return 1
print(f"superseded question {args.question_id}", file=out)
return 0
def _build_coordinator(args: argparse.Namespace) -> Any:
"""Construct a :class:`Coordinator` for the ``start`` / ``serve`` commands.
The live transport is built LAZILY here (never at import) so ``run-team.py``
imports, ``--help``, and the ledger subcommands all work with no Slack token
present. ``--dry-run`` substitutes a non-posting transport so an operator can
drive intake without a token (the ledger row is still written; only the
transport post is a no-op).
The coordinator is imported inside this function for the same reason: pulling
in the live runtime (and its optional SDK-adjacent deps) must not happen just
to render ``--help`` or run a read-only ledger command.
"""
from agent_team.coordinator import (
Coordinator,
default_plan_node_factory,
default_review_wiring,
)
transport = _build_transport(args)
# Production runs the full P2 graph: the wrapped real planner + the bound
# GPT-4.1 review loop (Plane-2 depth-first). These factories are lazy and
# only build/bind the model seams when a task actually runs.
return Coordinator(
db_path=args.db,
transport=transport,
build_plan_node=default_plan_node_factory,
review_wiring=default_review_wiring,
)
def _build_transport(args: argparse.Namespace) -> Any:
"""Build the transport for a coordinator command (lazy; token-tolerant).
``--dry-run`` (or any transport in dry-run) yields a non-posting transport so
intake works without credentials. Otherwise the live transport is built
lazily from the environment for the chosen ``--transport`` (so import,
``--help``, and ledger commands never need a token):
* ``slack`` -> :func:`build_live_slack_transport` over ``SLACK_BOT_TOKEN`` /
``SLACK_CHANNEL_ID``.
* ``github`` -> :func:`build_live_github_transport` over ``GITHUB_TOKEN`` and
the issue thread ``GITHUB_OWNER`` / ``GITHUB_REPO`` /
``GITHUB_ISSUE_NUMBER``. This is the §3.3.1 human-gate I/O only (post an
issue/PR comment); it carries NO CI, OIDC, git/patch apply, or GitHub
Actions network (the P3 apply/verify workflow is held for the security
gate).
* ``claude_code`` -> :func:`build_live_claude_code_transport` over the local
file-drop directory ``CLAUDE_CODE_DROP_DIR`` the Mac harness polls (D10).
Each live builder defers its optional SDK / token resolution to call time, so
a missing dependency or credential fails loudly here rather than at import.
"""
if getattr(args, "dry_run", False):
return _DryRunTransport()
if args.transport == "slack":
from agent_team.transport.slack_live import build_live_slack_transport
channel = os.environ.get("SLACK_CHANNEL_ID", "")
return build_live_slack_transport(channel)
if args.transport == "github":
from agent_team.transport.github_live import build_live_github_transport
owner, repo, issue_number = _github_thread_from_env()
# Token resolves from GITHUB_TOKEN inside the builder (call-time read);
# a missing token fails loudly there rather than being captured here.
return build_live_github_transport(
owner=owner,
repo=repo,
issue_number=issue_number,
token=os.environ.get("GITHUB_TOKEN") or None,
)
if args.transport == "claude_code":
from agent_team.transport.claude_code_live import (
build_live_claude_code_transport,
)
drop_dir = os.environ.get("CLAUDE_CODE_DROP_DIR", "")
if not drop_dir:
raise SystemExit(
"live transport 'claude_code' requires CLAUDE_CODE_DROP_DIR (the "
"Mac file-drop directory the Claude-Code harness polls); set it, "
"or use --dry-run for a no-token dry run"
)
return build_live_claude_code_transport(drop_dir)
raise SystemExit(
f"live transport '{args.transport}' is not wired for the run-team CLI; "
"use --transport slack/github/claude_code, or --dry-run for a no-token "
"dry run"
)
def _github_thread_from_env() -> tuple[str, str, int]:
"""Resolve the GitHub issue thread (owner/repo/issue) from the environment.
The live GitHub transport posts the clarifier question-set as a comment on a
fixed ``owner/repo#issue_number`` thread, so the thread is configured via
``GITHUB_OWNER`` / ``GITHUB_REPO`` / ``GITHUB_ISSUE_NUMBER`` (read lazily so
the value is never captured at import). A missing or non-integer value raises
a clear :class:`SystemExit` rather than building a half-configured transport.
"""
owner = os.environ.get("GITHUB_OWNER", "")
repo = os.environ.get("GITHUB_REPO", "")
raw_issue = os.environ.get("GITHUB_ISSUE_NUMBER", "")
missing = [
name
for name, value in (
("GITHUB_OWNER", owner),
("GITHUB_REPO", repo),
("GITHUB_ISSUE_NUMBER", raw_issue),
)
if not value
]
if missing:
raise SystemExit(
"live transport 'github' requires "
f"{', '.join(missing)}; set the issue thread (owner/repo/issue), or "
"use --dry-run for a no-token dry run"
)
try:
issue_number = int(raw_issue)
except ValueError:
raise SystemExit(
f"GITHUB_ISSUE_NUMBER must be an integer, got {raw_issue!r}"
) from None
return owner, repo, issue_number
class _DryRunTransport(Transport):
"""A non-posting transport for ``--dry-run`` intake (no token, no Slack).
``post_question`` records nothing on a real channel — it returns a synthetic
``channel_ref`` so :func:`agent_team.responder.notify_question` still writes
and stamps the durable ledger row (the durable seam is exercised; only the
side-effecting post is skipped). ``parse_answer`` is unused by the CLI path
but implemented so the ABC is concrete.
"""
def post_question(
self,
*,
thread_id: str,
question_id: str,
turn: int,
question_set: Any,
deadline: str,
) -> str:
print(
f"[dry-run] would post question {question_id} (turn {turn}) "
f"for thread {thread_id}",
file=sys.stderr,
)
return f"dry-run:{question_id}"
def parse_answer(self, raw: Any) -> tuple[str, Any, str]:
raise NotImplementedError("dry-run transport does not parse answers")
def _cmd_start(args: argparse.Namespace, *, out: Any) -> int:
"""Intake: start one task and run it to the first human gate (§3.3, §3.3.1).
Builds a :class:`Coordinator` (transport from the lazy factory; ``--dry-run``
posts nowhere), runs ``setup`` + ``start_task``, and prints the minted
``thread_id``. The clarifier question-set is delivered over the chosen
transport (or no-op under ``--dry-run``); the durable ledger row is written
either way.
"""
coordinator = _build_coordinator(args)
# start_task runs the clarifier graph to the first human gate IN THIS PROCESS,
# and the clarifier calls Claude (assess_confidence). The invoker is a
# process-local binding, so a running daemon does not help this CLI process —
# bind the real subscription invoker here, mirroring Coordinator.serve(), or
# start_task fails with "claude_invoke has no invoker bound".
from agent_team.invoker import bind_subscription_invoker
bind_subscription_invoker()
coordinator.setup()
thread_id = coordinator.start_task(
task_text=args.task, transport_name=args.transport
)
print(thread_id, file=out)
return 0
def _cmd_serve(args: argparse.Namespace, *, out: Any) -> int:
"""Run the coordinator daemon loop (binds the live invoker; §7.1 P1).
Delegates to :meth:`agent_team.coordinator.Coordinator.serve`, which binds
the real Claude subscription invoker, runs the startup recovery sweep, then
loops on the deadline cadence. The Slack inbound feed is the slack_listener's
job; this command owns the maintenance loop. Runs until interrupted.
"""
coordinator = _build_coordinator(args)
print("agent-team coordinator starting (Ctrl-C to stop)", file=out)
coordinator.serve()
return 0 # pragma: no cover - serve() loops until interrupted
def _cmd_intake_github(args: argparse.Namespace, *, out: Any) -> int:
"""Poll a repo for labeled issues and start one task per new issue (§3.3.1).
The GitHub-issue intake front door: builds a :class:`Coordinator` (transport
from the lazy factory; ``--dry-run`` posts nowhere), runs ``setup``, then
constructs a :class:`agent_team.transport.github_intake.GithubIntake` over a
read-only REST issue client
(:func:`~agent_team.transport.github_intake.build_default_issue_client`,
deferred-import, reads ``GITHUB_TOKEN`` at call time) and runs ONE
:meth:`~agent_team.transport.github_intake.GithubIntake.poll_once`. An
operator (or a cron) re-runs the command on a cadence; de-dup is DURABLE
(the ``ingested_issues`` ledger table), so re-running over a still-open
labeled issue never starts a second task — safe for a scheduled timer.
OPT-IN and INERT: this only reads labeled issues and calls the committed
coordinator intake entry — no CI, OIDC, git/patch apply, or GitHub-Actions
network. Owner/repo/label come from CLI flags; the token comes from
``GITHUB_TOKEN``. Prints the issue ids ingested on this pass (one per line).
"""
from agent_team.transport.github_intake import (
GithubIntake,
build_default_issue_client,
build_ledger_ingest_store,
)
coordinator = _build_coordinator(args)
coordinator.setup()
client = build_default_issue_client(owner=args.owner, repo=args.repo)
# Durable de-dup: a scheduled/cron intake runs as a fresh process each time,
# so the in-memory default would re-ingest every still-open labeled issue and
# spawn duplicate tasks. The ledger store keys on (source, issue_id) in the
# same SQLite DB the coordinator uses, making intake idempotent across runs.
store = build_ledger_ingest_store(
db_path=args.db, source=f"github:{args.owner}/{args.repo}"
)
intake = GithubIntake(
client=client,
coordinator=coordinator,
label=args.label,
store=store,
)
ingested = intake.poll_once()
for issue_id in ingested:
print(issue_id, file=out)
if not ingested:
print("github-intake: no new labeled issues to ingest", file=sys.stderr)
return 0
def _cmd_intake_checker(args: argparse.Namespace, *, out: Any) -> int:
"""Read checker reports and start one task per confirmed eligible finding.
The Plane-1 -> Plane-2 cross-plane front door (P5): builds a
:class:`Coordinator` (transport from the lazy factory; ``--dry-run`` posts
nowhere), runs ``setup``, then constructs a
:class:`agent_team.transport.checker_intake.CheckerFindingIntake` and runs
ONE :meth:`~agent_team.transport.checker_intake.CheckerFindingIntake.ingest_reports`
over the report paths. Each ``--report`` may be a checker report JSON file or
a directory of ``*.json`` reports.
OPT-IN and INERT: this only reads local report JSON and calls the committed
coordinator intake entry — no CI, OIDC, git/patch apply, or network. The
severity threshold (default ``high``) and transport come from CLI flags.
De-dup is in-memory per process, so each run is a single pass. Prints the
finding identities ingested on this pass (one per line).
"""
from agent_team.transport.checker_intake import CheckerFindingIntake
coordinator = _build_coordinator(args)
coordinator.setup()
intake = CheckerFindingIntake(
coordinator=coordinator,
threshold=args.threshold,
transport_name=args.transport,
)
ingested = intake.ingest_reports([Path(p) for p in args.report])
for identity in ingested:
print(identity, file=out)
if not ingested:
print(
"checker-intake: no new confirmed at/above-threshold findings to ingest",
file=sys.stderr,
)
return 0
def _load_finding_for_fix(args: argparse.Namespace) -> dict[str, Any]:
"""Load the single confirmed dependency-cve finding to fix from a report file.
Reads the ``dependency-cve.json`` report (the checker's output shape) at
``--report`` and selects the finding by ``--finding-id``. Returns the finding
mapping. Raises :class:`SystemExit` on any load/selection error so the CLI
fails loudly rather than dispatching against a half-resolved finding. Pure
read; no network, no CI, no git.
"""
report_path = Path(args.report)
try:
raw = report_path.read_text(encoding="utf-8")
except OSError as exc:
raise SystemExit(f"cannot read finding report {report_path}: {exc}") from exc
try:
report = json.loads(raw)
except ValueError as exc:
raise SystemExit(
f"finding report {report_path} is not valid JSON: {exc}"
) from exc
findings = report.get("findings") if isinstance(report, dict) else None
if not isinstance(findings, list):
raise SystemExit(
f"finding report {report_path} has no 'findings' array (got "
f"{type(report).__name__})"
)
matches = [
f for f in findings if isinstance(f, dict) and f.get("id") == args.finding_id
]
if not matches:
raise SystemExit(f"no finding with id {args.finding_id!r} in {report_path}")
if len(matches) > 1:
raise SystemExit(
f"ambiguous: {len(matches)} findings share id {args.finding_id!r} in "
f"{report_path}"
)
return matches[0]
def _cmd_fix(args: argparse.Namespace, *, out: Any) -> int:
"""Plan a Plane-1 dep-bump fix and show what it WOULD dispatch (§7 Phase 5).
The Tier-3 fixer front door (design §4 fixer row, §3.3.2). Loads ONE confirmed
``dependency-cve`` finding from the report, asks the fixer to produce a fix
spec (Claude) + a minimal bump patch (DeepSeek via the orchestrator) + the CI
``workflow_dispatch`` inputs, and — in ``--dry-run`` (the only mode wired
here) — PRINTS the spec, patch, and dispatch inputs without dispatching
anything.
OPT-IN / INERT: this command never dispatches. It binds NO workflow dispatcher
(the box holds no write token, D2), so even an ``ok`` plan only prints. Live
dispatch is a provisioning-time wiring of the trusted apply path's dispatcher,
deliberately not reachable from this CLI. A non-fixable finding prints the
fail-safe reason and exits non-zero.
Returns ``0`` when a fix plan was produced (dry-run printed), ``1`` when the
finding is not fixable (fail-safe; nothing planned).
"""
from agent_team.nodes.fixer import describe_plan, plan_fix
if not args.dry_run:
# Live dispatch is provisioning-gated and not wired into the CLI; refuse
# to run without --dry-run rather than silently doing nothing.
raise SystemExit(
"fix supports only --dry-run in this build (live dispatch is "
"provisioning-gated; the box holds no write token, D2). Re-run with "
"--dry-run to see what it WOULD dispatch."
)
finding = _load_finding_for_fix(args)
plan = plan_fix(finding, task_id=args.task_id)
print(describe_plan(plan), file=out)
return 0 if plan.ok else 1
def _cmd_force_resume(args: argparse.Namespace, *, out: Any) -> int:
"""Force-resume a parked task's question (destructive; audit-logged).
The design-named operator verb (§3.3.1 / §6.6 "an operator can force-resume
... a parked task via the CLI"). A task parks when its clarifier question
EXPIRES with no answer, so the un-park action is to RE-OPEN that expired
question (:func:`agent_team.db.schema.reopen_question`) so the normal
delivery → answer → resume flow can proceed.
Crucially this does NOT ``supersede`` the row: superseding an ``answered``
row would flip it out of the state the recovery sweep resumes from, making a
stuck-but-answered task permanently un-resumable — the opposite of
force-resume. So:
* ``expired`` (the parked case) → reopened; returns 0.
* ``answered`` (answered but not yet resumed) → already eligible for the
recovery resume sweep; intent is recorded and we report that, no mutation.
* ``open`` / ``superseded`` / absent → nothing to force; reported as a no-op.
"""
_require_confirm("force-resume", confirm=args.confirm)
_audit_attempt(
args.audit_log,
"force-resume",
question_id=args.question_id,
operator=args.operator,
detail={"resume_requested": True},
)
conn = connect(args.db)
try:
row = _fetch_question(conn, args.question_id)
status = None if row is None else row["status"]
reopened = False
if status == "expired":
reopened = reopen_question(conn, question_id=args.question_id)
finally:
conn.close()
_audit_outcome(
args.audit_log,
"force-resume",
question_id=args.question_id,
operator=args.operator,
applied=reopened,
detail={"resume_requested": True, "prior_status": status},
)
if reopened:
print(
f"force-resume: reopened expired question {args.question_id}; "
"it will be re-delivered for an answer",
file=out,
)
return 0
if status == "answered":
print(
f"force-resume: question {args.question_id} is answered and pending "
"resume; the recovery sweep will resume it (intent recorded)",
file=out,
)
return 0
print(
f"force-resume no-op: question {args.question_id} "
f"({'absent' if status is None else f'status={status}'}) is not parked",
file=sys.stderr,
)
return 1
# Statuses an operator treats as "parked context": a task whose only pending
# question is no longer open may be parked (answered-but-unresumed, expired, or
# superseded). ``open`` is excluded — that is the live-waiting view (default
# ``list``). Derived from the foundation QUESTION_STATES so it stays in sync.
_PARKED_STATES: tuple[str, ...] = tuple(s for s in QUESTION_STATES if s != "open")
def build_parser() -> argparse.ArgumentParser:
"""Construct the argparse parser for ``run-team.py`` (no side effects)."""
parser = argparse.ArgumentParser(
prog="run-team.py",
description=(
"R720 agent-team operator CLI — manual path over the durable "
"pending_questions ledger (design §3.3.1)."
),
)
parser.add_argument(
"--db",
type=Path,
default=_DEFAULT_DB,
help=f"path to the agent-team SQLite ledger (default: {_DEFAULT_DB})",
)
parser.add_argument(
"--audit-log",
type=Path,
default=_DEFAULT_AUDIT_LOG,
dest="audit_log",
help=(
"append-only JSONL audit log for destructive actions "
f"(default: {_DEFAULT_AUDIT_LOG})"
),
)
parser.add_argument(
"--operator",
default=_default_operator(),
help="operator identity recorded in the audit log for destructive actions "
"(defaults to the OS login so the trail is always attributable)",
)
sub = parser.add_subparsers(dest="command", required=True)
p_init = sub.add_parser("init-db", help="create/upgrade the ledger tables")
p_init.set_defaults(func=_cmd_init_db)
p_list = sub.add_parser("list", help="list pending questions (read-only)")
list_filter = p_list.add_mutually_exclusive_group()
list_filter.add_argument(
"--status",
choices=QUESTION_STATES,
default="open",
help="lifecycle status to list (default: open)",
)
list_filter.add_argument(
"--all",
action="store_true",
help="list questions in every lifecycle status",
)
list_filter.add_argument(
"--parked",
action="store_true",
help="list non-open questions (parked-task context)",
)
p_list.set_defaults(func=_cmd_list)
p_show = sub.add_parser("show", help="print one question row (read-only)")
p_show.add_argument("question_id", help="the question_id to show")
p_show.set_defaults(func=_cmd_show)
p_redeliver = sub.add_parser(
"redeliver",
help="clear an open question's channel_ref so it is re-posted",
)
p_redeliver.add_argument("question_id", help="the question_id to re-deliver")
p_redeliver.set_defaults(func=_cmd_redeliver)
p_expire = sub.add_parser(
"expire", help="force-expire an open question (destructive)"
)
p_expire.add_argument("question_id", help="the question_id to expire")
p_expire.add_argument(
"--confirm",
action="store_true",
help="required: confirm this destructive, audit-logged action",
)
p_expire.set_defaults(func=_cmd_expire)
p_answer = sub.add_parser(
"answer", help="answer a question on a task's behalf (destructive)"
)
p_answer.add_argument("question_id", help="the question_id to answer")
p_answer.add_argument(
"--answer",
required=True,
help="the answer payload (stored verbatim as answer_json)",
)
p_answer.add_argument(
"--via",
default="",
help="answering identity for answered_via (default: cli:<operator>)",
)
p_answer.add_argument(
"--confirm",
action="store_true",
help="required: confirm this destructive, audit-logged action",
)
p_answer.set_defaults(func=_cmd_answer)
p_supersede = sub.add_parser(
"supersede", help="mark a stale question superseded (destructive)"
)
p_supersede.add_argument("question_id", help="the question_id to supersede")
p_supersede.add_argument(
"--confirm",
action="store_true",
help="required: confirm this destructive, audit-logged action",
)
p_supersede.set_defaults(func=_cmd_supersede)
p_resume = sub.add_parser(
"force-resume",
help="force-resume a parked task's question (destructive)",
)
p_resume.add_argument("question_id", help="the question_id to force-resume")
p_resume.add_argument(
"--confirm",
action="store_true",
help="required: confirm this destructive, audit-logged action",
)
p_resume.set_defaults(func=_cmd_force_resume)
p_start = sub.add_parser(
"start",
help="intake: start one task and run it to the first human gate",
)
p_start.add_argument(
"--task",
required=True,
help="the task description (intake text) to run through the pipeline",
)
p_start.add_argument(
"--transport",
choices=_TRANSPORT_CHOICES,
default="slack",
help="channel for delivering clarifier questions (default: slack)",
)
p_start.add_argument(
"--dry-run",
action="store_true",
dest="dry_run",
help="use a non-posting transport (no token needed; ledger still written)",
)
p_start.set_defaults(func=_cmd_start)
p_serve = sub.add_parser(
"serve",
help="run the coordinator daemon (binds invoker, runs the loop)",
)
p_serve.add_argument(
"--transport",
choices=_TRANSPORT_CHOICES,
default="slack",
help="channel for delivering clarifier questions (default: slack)",
)
p_serve.add_argument(
"--dry-run",
action="store_true",
dest="dry_run",
help="use a non-posting transport (no token needed)",
)
p_serve.set_defaults(func=_cmd_serve)
p_intake = sub.add_parser(
"intake-github",
help="poll a repo for labeled issues and start one task per new issue",
)
p_intake.add_argument(
"--owner",
required=True,
help="GitHub repository owner / org login to poll for intake issues",
)
p_intake.add_argument(
"--repo",
required=True,
help="GitHub repository name to poll for intake issues",
)
p_intake.add_argument(
"--label",
required=True,
help="issue label that flags an issue as pipeline intake (non-empty)",
)
p_intake.add_argument(
"--transport",
choices=_TRANSPORT_CHOICES,
default="github",
help="channel for delivering clarifier questions (default: github)",
)
p_intake.add_argument(
"--dry-run",
action="store_true",
dest="dry_run",
help="use a non-posting transport (no token needed; ingest still runs)",
)
p_intake.set_defaults(func=_cmd_intake_github)
p_intake_checker = sub.add_parser(
"intake-checker",
help=(
"read checker reports and start one remediation task per confirmed "
"at/above-threshold finding (P5 cross-plane loop)"
),
)
p_intake_checker.add_argument(
"--report",
required=True,
action="append",
metavar="PATH",
help=(
"checker report JSON file or directory of *.json reports; repeatable "
"to ingest several reports in one pass"
),
)
p_intake_checker.add_argument(
"--threshold",
default="high",
choices=("low", "medium", "high", "critical"),
help=(
"minimum severity a confirmed finding must reach to spawn a task "
"(default: high)"
),
)
p_intake_checker.add_argument(
"--transport",
choices=_TRANSPORT_CHOICES,
default="github",
help="channel for delivering clarifier questions (default: github)",
)
p_intake_checker.add_argument(
"--dry-run",
action="store_true",
dest="dry_run",
help="use a non-posting transport (no token needed; ingest still runs)",
)
p_intake_checker.set_defaults(func=_cmd_intake_checker)
p_fix = sub.add_parser(
"fix",
help=(
"plan a Plane-1 dep-bump fix for a confirmed dependency-cve finding "
"and (dry-run) show what it WOULD dispatch to org CI"
),
)
p_fix.add_argument(
"--report",
required=True,
help="path to the dependency-cve.json finding report to read the finding from",
)
p_fix.add_argument(
"--finding-id",
required=True,
dest="finding_id",
help="the finding id (report findings[].id) to fix",
)
p_fix.add_argument(
"--task-id",
required=True,
dest="task_id",
help="pipeline task id (provenance; becomes the CI dispatch task_id)",
)
p_fix.add_argument(
"--dry-run",
action="store_true",
dest="dry_run",
help=(
"show the spec + patch + dispatch inputs WITHOUT dispatching "
"(the only supported mode; live dispatch is provisioning-gated)"
),
)
p_fix.set_defaults(func=_cmd_fix)
return parser
def main(argv: Sequence[str] | None = None, *, out: Any = None) -> int:
"""CLI entry point. Returns a process exit code.
``argv`` defaults to ``sys.argv[1:]``; ``out`` defaults to ``sys.stdout``
(injectable for tests). Operational failures return ``1``; a missing
``--confirm`` on a destructive action raises :class:`PermissionError`,
surfaced as exit code ``1`` with a stderr message.
"""
out = out if out is not None else sys.stdout
parser = build_parser()
args = parser.parse_args(argv)
try:
return int(args.func(args, out=out))
except PermissionError as exc:
# A refused destructive action (no --confirm) or an unwritable audit
# path. The attempt-before-mutate ordering means nothing was mutated.
print(f"error: {exc}", file=sys.stderr)
return 1
except OSError as exc:
# Any other audit-log / filesystem failure (e.g. the audit append could
# not be written). Surfaced cleanly instead of as an uncaught traceback;
# if the attempt record was written, the action is on the trail.
print(f"error: audit/IO failure: {exc}", file=sys.stderr)
return 1
if __name__ == "__main__": # pragma: no cover
raise SystemExit(main())