Repository navigation
feat(asks): agents withdraw their own pending ask — withdraw route, withdraw_ask tool, Workspace/Operating Room history (trinity-enterprise#901) - #3728
Merged
Conversation
…enterprise#901)
Checkpoint A of the withdraw-own-ask feature: an agent ends one of its OWN
pending asks with a reason and no successor.
- db: `_replace_predecessor` is generalised into `_end_own_pending(conn,
agent, row_id, *, reason, replaced_by, honour_deadline)` — the one
compare-and-set an agent ends its own ask with (ownership belt:
agent_name + raised_by='agent' on the column + status='pending'). Replace
keeps its vocabulary and its deadline clause (T5b expiry stays); the new
`withdraw_item` runs it without the deadline clause, like a person's
`cancel_item` (a withdraw lands no approval, so nothing is denied by
timeout and the C6 re-ask guard is not tripped). Writes
cancelled / cancelled / disposed_by='agent' / disposition_reason=<reason>
/ replaced_by NULL. Outcomes: withdrawn | ended | not_own.
- service: `WITHDRAWN` beside `REPLACED`; `agent_ending_kind(row)` tells the
two apart ONCE, by the link (never by the reason's text), and rides every
receipt and the readback as the explicit `agent_ending` field.
`ask_service.withdraw()` — reason whitespace-collapsed and bounded at
ingest (422 invalid_reason); uniform 404 ask_not_found for an unknown /
another agent's / gate / platform / about-a-person row; 409 ask_ended
with ids and enums only when a person answered or ended it first; 200
replayed on a repeat (a changed reason reported in `differs`, never
applied). One ending sink: an agent-keyed `withdrawn` audit row carrying
has_reason only (never the text), the thin operator_queue_cancelled
trigger, no wake (_wake_filer skips agent-authored endings).
`_refuse_ended_predecessor` now wraps the shared `_refuse_ended`, which
gains a withdrawn arm for a replace of a withdrawn predecessor.
- route: POST /api/agents/{name}/operator-queue/{request_id}/withdraw
behind get_self_acting_agent (agent-only by design;
reject_non_person_principal deliberately not applied), AskRejected
mapped to {code, message, ...}; listed as agent-callable in the #2996
route census.
- tests: tests/unit/test_ent901_withdraw_ask.py (59, written red first —
58 failed on the missing behaviour, the replace-regression pin was the
one green). Deliberate re-pins: RECEIPT_KEYS in test_3247 and
test_ent611_native_ask, READBACK_KEYS in test_ent611_ask_endings gain
`agent_ending`.
Not in this commit (later checkpoints): the MCP tool `withdraw_ask`, the
prompt twins + sentinels, the Workspace projection, the Inbox row, the
Operating Room card reason, docs and the router `# mcp:` header.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…se#901)
Checkpoint B of trinity-enterprise#901 — the agent-facing surface over the
withdraw route checkpoint A shipped.
- MCP tool `withdraw_ask(request_id, reason)` in tools/operator_queue.ts,
self-acting (resolveActingAgent; access.ts `kind: "none"`). Only
`{reason}` travels in the body; refusals (404 ask_not_found, 409
ask_ended, 422 invalid_reason) come back through askRefusal as their
named code with extras, never a throw. client.withdrawAsk, the
OperatorAskWithdraw type, `agent_ending` on the receipt/readback types,
and "withdrawn" on the receipt status union.
- Descriptions: ask_operator gains "if nothing is needed any more,
withdraw_ask" in its replaces sentence (kept under the 1,800 cap by
dropping "at once"); get_my_ask names a withdrawn ending and
`agent_ending`; respond_to_operator_queue names withdraw_ask beside the
replace.
- Prompt contract: the "Withdraw what is moot." paragraph, byte-identical
in PLATFORM_INSTRUCTIONS and config/trinity-meta-prompt/prompt.md,
directly after the "Do not re-ask what is still pending" paragraph
(that line untouched). The Pending asks line offers "if it is moot,
withdraw it (withdraw_ask)" only when it carries an ask the agent raised.
- Router `# mcp:` header names withdraw_ask.
Tests (written red first): operator_queue.test.ts withdraw_ask block
(stray-keys execute, scope refusals, the three refusal codes, the
description pins); access-wiring.test.ts real-transport block (agent key
acts as itself, only {reason} reaches the backend, user key refused before
any call, 409 extras); test_1402 sentinels `withdraw_ask` and `ask_ended`;
test_ent901 TestPromptContract + TestPendingLineOffersWithdraw. access.test
went red on the missing TOOL_ACCESS_POLICY row before the entry.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…nd projection + docs (trinity-enterprise#901) Checkpoint C1 of trinity-enterprise#901 (C split: C1 backend projection + docs, C2 frontend + specs). - client_portal/asks/service.py: `_ending_of` keys its agent arm on `ask_service.agent_ending_kind` (replaced | withdrawn), still before the email arm; `_project` carries `agent_ending` for an agent ending and `ending_reason` (the agent's reason) for a WITHDRAW only — the operator's cancel reason, a platform token and a timeout never cross. `answer_ask`'s lost-race message says the agent withdrew it. - WorkspaceAsk gains `agent_ending` / `ending_reason`. - Operator person view (`_for_principal`) carries `agent_ending`. - Execution Context Ended asks line: `list_recent_endings_for_agent` selects `replaced_by` (still ids + enums only) and `_render_ended_asks` says `withdrawn` vs `replaced` via `agent_ending_kind` — never the reason. - Docs: requirements/security.md §26.9 / §26.10, architecture/ api-endpoints.md, feature-flows/operating-room.md; the endings-query and EndingEvent.reason docstrings. Tests (red first, 9 new in test_ent901_withdraw_ask.py): projection kind + reason, a reason literally `replaced` stays a withdraw, a replace carries no reason, no other ending forwards a reason, the withdrawn ask leaves the open list and shows as history, the lost-race message, the Ended asks line (withdrawn, reason absent) and its select set, the operator person view. Mutation: dropping the withdrawn arm in `_ending_of` turns 3 projection tests red. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…end + specs (trinity-enterprise#901) Checkpoint C2 of trinity-enterprise#901. - utils/operatorQueue.js: `endingKind` reads the server's derived `agent_ending` first (replaced | withdrawn), falling back to `replaced` for an agent ending on an older payload; `ENDING_LABELS.withdrawn`; `queueEnding` gives a withdraw `who: 'the agent'` and the agent's reason (`ending_reason` on the Workspace, `disposition_reason` on an operator row) as `reason`, kept out of the ending's words, so `workspaceEndingText` reads "Withdrawn by the agent". - PortalInboxList: the ended badge stays "Withdrawn by the agent"; the reason is its own single-line clamped line with the full text on `title`. Never a Waiting item (it leaves openAsks by status). - PortalAsks: the ending line reads "Withdrawn by the agent · <when>" with the reason on its own clamped line. - ResolvedCard (T3): the note shows the withdraw reason; a replace stays blank (its reason is the closed token `replaced`). Text interpolation only on every surface; no new palette classes. Specs (red first): operatorQueueWithdraw.spec.js (the rule, incl. a reason literally "replaced"), portalInboxWithdrawn.mount.spec.js (jsdom mount of the Inbox row, PortalAsks, openAsks / count), and resolvedCardWithdrawn.spec.js (mounted note; replace has none). Mutation: removing the `agent_ending` read in `endingKind` turns 6 of the 13 red. Full frontend unit suite: 368 files / 5,913 tests pass. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…oc lists withdraw_ask (trinity-enterprise#901) The ent329 CAS-win guard found `_ended` feeders by the text `_ended(`, which also matched `_refuse_ended(` / `_refuse_ended_predecessor(` (helpers that end nothing). Detection is now an AST walk for real `_ended(...)` calls, and the call-position index uses a regex that excludes identifiers ending in `_ended(`. `withdraw` is guarded by its db writer's `outcome`, a form the guard did not know. It now accepts an outcome-form CAS writer only when the ending function refuses every outcome but the pinned winning token (`!= "withdrawn"` → raise/return) before `_ended`; a deny-list of the losing outcomes is not accepted. `withdraw` gains that two-line positive check (fail closed on an unknown outcome), so it no longer relies on a deny-list. `withdraw` is now in the guard's required-feeder floor. MCP docs: operator_queue.ts lists six tools, including withdraw_ask. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Contributor
Author
|
Merge-tree record (orchestrator, 2026-10-14): this branch merges cleanly onto live |
…nterprise#901) The person view of the operator queue gained the derived agent_ending enum (replaced | withdrawn | None) in checkpoint A; ent#715's classification guard pins the person key set and named it as unplaced. It sits beside the other person-only derived keys (decided_by_options, gate): never a column, never PII. The machine allowlist is unchanged — derived keys are person-only there, and the agent reads agent_ending on its own readback. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
trinity-ability
approved these changes
Oct 14, 2026
trinity-ability
left a comment
trinity-ability
left a comment
Contributor
There was a problem hiding this comment.
/validate-pr — APPROVE (Lane B, P1 feature, no schema change)
Validated at head f9c7cf459 against dev fa6c9df56. git merge-tree against live dev is clean.
| Check | Result |
|---|---|
| Base / size | dev · 36 files (+1899/−74) |
| Issue link | Fixes Abilityai/trinity-enterprise#901 resolves (P1, type-feature). Cross-tracker, so set status-in-dev on ent#901 by hand after merge and close it by hand at release |
| Commits | 6 commits, all conventional and descriptive |
| Docs | requirements security.md §26.9/§26.10, architecture/api-endpoints.md, architecture/mcp-server.md, feature-flows/operating-room.md, user-docs all updated. The enterprise-docs guard pattern finds nothing in the added lines |
| Security §4.1–4.10 | No keys, emails, IPs, env files, secrets, symlinks or host paths. No new env vars and no new top-level backend package |
| Auth | New route is behind get_self_acting_agent and is in the route census as agent-callable. Ownership is checked twice: the service gate (raised_by == 'agent', is_about_a_person) and the CAS predicate. The skill-gate ending observer is unaffected because gate rows are a uniform 404 |
Tests in tests/ with deleted lines |
The ent#329 CAS guard got stricter: 5 required feeders, it parses real ast.Calls, and the outcome form is accepted only on an exact winning-token != check. The other two files only add agent_ending to key sets |
| Merge gate (Tier 1) | All green, including pytest (head) + regression diff. dev-ci on dev is success and there is no open dev-ci red issue |
Execution coverage (§5.4). Run in the PR worktree.
test_ent901_withdraw_ask.py: 74/74. With the ent329 and 3247 neighbours: 197/197.- Mutations (each run on its own, reverted after):
- drop
status == 'pending'from the CAS: 20 red - drop the
raised_bygate in both layers: 4 red agent_ending_kindalways returnswithdrawn: 8 red- withdraw honours the deadline: 2 red
- remove the
outcome != "withdrawn"fail-closed guard: 1 red (the ent329 guard)
- drop
- Route tests drive the real
agent_routerwith the realget_self_acting_agent. Onlyget_current_useris overridden. Negative principals (a sibling agent's key, a person, a user key, a system key, a connector) each get a 403 and the row stays pending. - Frontend: the 3 new specs are jsdom
mount()s ofResolvedCard,PortalInboxListandPortalAsks, not source-text checks. Together with the source-text, raw-color and loading-gate ratchets: 41/41. - MCP:
tsc --noEmitclean,npm test952/952.withdraw_askis registered throughcreateOperatorQueueToolsandaccess.ts,request_idis URL-encoded, and the real-transport wiring test is included.
Notes (not blocking)
- The structural
/review(MERGEABLE) and/cso --diff(0 findings) are recorded in the orchestrator's plan files, not as a review on this PR. - The cso hypothesis still applies:
EndingEvent.reasonnow has two authors, so a future observer that forwards it must branch ondisposed_by. - The operator owns the Before merge items: rc UI check, MCP server restart, stale tool counts. This review does not claim them.
vybe
marked this pull request as ready for review
October 14, 2026 19:01
Contributor
Author
|
merge-train (2026-10-14): merged at f9c7cf4 after a fresh structural /review in an isolated worktree — READY, 0 critical. 607 pytest / 952 MCP / 21 vitest green; 16 fix mutations each turned a test red (CAS pending-only, raised_by gate, is_about_a_person, router decorator, include_router, prompt paragraph, frontend spans). Not blocking, recorded for the author:
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Fixes Abilityai/trinity-enterprise#901 — an agent can end its own pending ask without raising a new one. Independent draft off
dev(chain-asks-1014, trinity-pm orchestrator, 2026-10-14). Draft until the operator merges.What
services/ask_service.py,db/operator_queue.py,database.py): one agent-side ending door,withdraw(agent, request_id, reason). The replace path's writer is generalised into_end_own_pending(compare-and-set), so withdraw and replace share one writer and one ending sink. A withdrawn row endscancelledbyagent,replaced_byempty, reason indisposition_reason, deadline clause ignored as for a person's cancel, no wake, audit row carrieshas_reasononly. Refusals: uniform 404ask_not_found(missing, another agent's, gate-raised, platform, queue-file rows), 409ask_endedwhen the person answered first (the answer stands), 200replayedon a repeat by the same agent, 422invalid_reason(strip + cap 1–500 at ingest). A positiveoutcome == "withdrawn"check sits before the ending sink. No Alembic revision.routers/operator_queue.py):POST /api/agents/{name}/operator-queue/{request_id}/withdrawbehindget_self_acting_agent; route census lists it as agent-callable.agent_ending: the kind of an agent ending (replaced/withdrawn) is derived once server-side fromreplaced_by, never from reason text, and projected explicitly on the receipt,get_my_ask, the operator readback and the Workspace — a replaced ask whose successor is withheld from the viewer can never render as withdrawn.src/mcp-server): new toolwithdraw_ask(request_id, reason)(access.tskindnone, sends only{reason});ask_operator/get_my_ask/respond_to_operator_queuedescriptions name it;ask_operatordescription stays ≤1,800 chars (dropped the words "at once", no documented behaviour changed).platform_prompt_service.py::PLATFORM_INSTRUCTIONSandconfig/trinity-meta-prompt/prompt.mdright after "Do not re-ask what is still pending" (that line untouched);_render_pending_asksadvice; test_1402 sentinelswithdraw_ask/ask_ended.client_portal/asks/service.py,utils/operatorQueue.js,PortalInboxList.vue,PortalAsks.vue,ResolvedCard.vue): a withdrawn ask leaves the Action list by status and shows as history "Withdrawn by the agent" with the reason on its own clamped line, never a Waiting item;ending_reasonis forwarded only for an agent withdraw; the Operating Room resolved card shows the reason for a withdraw and nothing for a replace; the Execution Context's ended-asks line says withdrawn and never the reason;answer_askconflict message names the withdraw.requirements/security.md§26.9/§26.10,architecture/api-endpoints.md,architecture/mcp-server.md,feature-flows/operating-room.md,user-docs/integrations/mcp-server.md, docstrings.Rulings carried (orchestrator, on the operator's behalf — plan file)
T1 a new MCP tool rather than a mode on
ask_operator;replaceswithout a successor stays rejected · T2 no new revision, reason indisposition_reason, kind projected as a field · T3 the Operating Room card shows the withdraw reason · T4 a repeated withdraw answers 200replayed· T5 no new Resolved-tab chip. All five follow the planner's recommendation.Review + security
/review(Fable, reading): MERGEABLE, 0 critical; informational I1 (MCP area doc count, fixed in0f34d8088) and I2 (machine-key operator view omitsagent_ending— deliberate allowlist, noted)./cso --diff: 0 findings; one hypothesis recorded —EndingEvent.reasonnow has two authors, so a future observer that forwards it must branch ondisposed_by. Evidence sweep (Opus, pytest-randomly, 20 files one per process): 19 green, 1 red branch-introduced — the ent#329 static CAS guard matched the new_refuse_ended*helper names and did not know the outcome form. Fixed in0f34d8088: the guard parses real_endedcalls and accepts the outcome form by a named token; mutation proof (outcome check removed) goes red namingwithdraw. CI round 1 on0f34d8088:regression diffnamed six parametrisations oftest_ent715_queue_person_fields::test_a_person_reads_every_person_field(the person key set is a literal in that test;agent_endingwas unlisted). Fixed inf9c7cf459, test-only: the derived person-only key set gainsagent_ending; the machine allowlist stays as it was.Tests
Backend: new
test_ent901_withdraw_ask74 (red-first); related ask files 925 → 715 → 361 across checkpoints, 512 in one process with thesys.modules-stubbing neighbours. MCP: 952/952 undernode --test,tsc --noEmitclean. Frontend: 3 new specs 13/13 (jsdom mount), related 303, fullnpm run test:unit5,913. Not run here: the full backend island (CI), live stack.Before merge
operatorhas no Workspace addressee at all.withdraw_askand the contract paragraph on their next turn (the platform prompt is read per turn); the MCP server needs a rebuild/restart for the new tool.user-docs/integrations/mcp-server.mdopening line,CLAUDE.md,AGENTS.md— their own fix.devand the sibling scan: first comment below.Handoffs
replacesas a close-out once this ships (rule already recorded on its side).🤖 Generated with Claude Code