Skip to content

feat(access): hold an agent's turn to what its person may do, and declare every agent route's tier (ent#926) - #3778

Merged
dolho merged 48 commits into
devfrom
feature/ent926-route-census
Nov 3, 2026
Merged

dolho merged 48 commits into
devfrom
feature/ent926-route-census

Conversation

@dolho

@dolho dolho commented Oct 16, 2026 •

Copy link
Copy Markdown
Contributor

Summary

An agent acts for whoever is talking to it. This PR makes what the agent may do in a turn follow who that turn is for, decided by the platform from the execution row, and makes every route that names an agent declare who may call it. Rationale, rulings and the threat detail live in abilityai/trinity-enterprise#926 (private tracker).

Post-1.0 item: do not merge before the 1.0 cut (operator ruling 2026-10-13 for the sibling items). Kept as a draft.

Stacked on #3776 and #3777. The shared path policy comes from #3776 and the per-person execution views from #3777. Until both land, this diff also shows their commits. Review order: #3776 → #3777 → #3778.

Who a turn is for

services/turn_tier_service.turn_actor decides from the execution row:

  • owner: the agent's owner, an admin, or something set up at owner tier. A schedule's runs (cron, webhook, retry, validation) are rated by the schedule's creator.
  • shared person: someone the agent is shared with.
  • visitor: a turn with no platform user, such as Slack, Telegram, WhatsApp, a public link, voice or a paid call.

The answer is served to the agent's own key at GET /api/agents/{name}/executions/{id}/turn-tier, as {tier, kind, shell}.

What changes at runtime

In the agent image

  • hooks/turn-tier-guard.py. It has its own root-owned managed-settings.d drop-in.
    • File tools: a shared person's or visitor's turn is held to the paths the Files API allows them.
    • Shell: owner-only. A visitor keeps it while the owner leaves the new guardrail on.
    • Paths outside the home are now classed: /app as runs-code, /var/lib/trinity/* as credential, /logs/* as system.
  • agent_server/services/turn_scope.py. Gemini CLI and Codex run no Claude Code hooks, so they refuse a turn they may not run, before the CLI starts. The refusal is a typed 403 with error_code: turn_refused, checked at spawn, so the push, sync chat and pull-mode paths all pass through it.

In the backend

  • Standing-work fence. An agent key acting in a turn that may not use the shell cannot create or change schedules (or their webhooks and triggers), loops, reminders, events or event subscriptions. The check sits at the agent-key auth entry point. POST /api/agents/{name}/emit-event is owner tier for people (agent keys keep the bug: loops, schedule triggers and agent events start a fresh depth-0 chain — an agent can launder its inter-agent chain depth (#2806 follow-up) #2973 sibling-emit path, fenced in restricted turns).
  • Dispatched turns carry their parent's person. A task, chat or fan-out the agent starts carries the person its parent turn is for. One helper sits at every stamping site, and an AST guard fails on a new source_user_id=current_user.id.
  • Narrowed key reads. In a shared person's turn, the agent's own key reads only that person's executions; this extends fix(access): scope per-user views of agent activity (ent#875) #3777's viewer tier.
  • A refusal is a 403. turn_refused is answered 403 on /chat, /task, the pull sink and the Workspace portal, never a retryable 503. A public-link visitor gets a plain line on both the sync path and the async status poll.

In the MCP server

  • Every backend call carries the tool call's X-Trinity-Execution-Id: an AsyncLocalStorage entered in withAudit, plus a single turnFetch door in client.ts. A guard fails on a bare fetch( there.

In the frontend

  • Settings › Guardrails gains "Channel and public visitors may run commands" (visitor_shell). It defaults on, so channel agents keep working, and it is read live per turn, with no restart. Its text states the trade-off.

Route census

Verification

Unit

  • test_ent926_turn_tier.py, test_ent926_turn_tier_hook.py, test_ent926_runtime_turn_scope.py, test_ent926_agent_key_turn_scope.py, test_ent926_agent_route_census.py and test_ent926_public_turn_refused.py.
  • Updated existing tests: the Gemini and Codex runtimes, the Files tiers, protected paths, and the ent#629 reach guard.
  • Mutation-checked: disabling the fence, reverting one stamping site, putting self_task back on the owner list, or removing the schedule-creator lookup each turns tests red.

Other suites

  • MCP: 970/970, including turn-context.test.ts over real HTTP.
  • Frontend: guardrailsVisitorShell.spec.js, plus both ratchets.

Live, on a local stack (2026-10-19). Fresh agents: Claude, Gemini, Codex, and a Claude and a Gemini pull pilot.

  • 63/63 deterministic API checks: route auth, Files tiers, viewer tiers, the fence on 12 routes × 6 turn kinds, dispatched-turn stamping, a webhook-fired run, and the agent key's reads.
  • 44/44 in-container hook checks: 11 paths × shared person / owner / visitor / no turn, with the switch on and off.
  • End-to-end turns through the model: shell, MCP reminders, delegation and fan-out.
  • All three non-Claude runtimes refuse a shared person with 403.
  • Two issues the live run found were fixed here: the MCP fan-out header, and the public visitor's refusal text.

Not covered live: the Workspace portal path (unit-tested), and a Codex owner turn past the access check.

Notes for the reviewer

  • The visitor switch defaults on, deliberately. Turning it off is the owner's choice. While on, a visitor's shell reaches what the agent can; the setting's text says so.
  • Not changed: the 115 legacy rulings. They follow as small PRs once the private ruling table is applied.

🤖 Generated with Claude Code

dolho and others added 4 commits October 16, 2026 12:51
The six agent Files routes inherited the accessor tier (admin | owner |
shared): anyone an agent was shared with could change its instructions and
startup files, list hidden directories, and act through links. Each verb now
resolves through one path policy, decided per tier (owner, share, key).

- services/file_path_policy.py: path classes, the tier x verb table and the
  agent server's platform floor; pure stdlib, vendored byte-identical into
  docker/base-image/agent_server/ (Invariant #5) with a parity test.
- Below the owner tier: instruction files, .claude/** (skills included) and
  files that run code are owner tier; the list withholds hidden files and
  flags each node with the verbs the caller cannot use; a file with a second
  hard-linked name is not served.
- For every caller: nested and bare-repo git metadata is write-protected;
  CLAUDE.md is not deletable; PUT/mkdir/DELETE and the list root refuse a link
  that changes any tier's verdict, then act through an O_NOFOLLOW dir_fd walk;
  DELETE of a link unlinks the link; the list never descends a link (20 000
  node cap); list/delete containment is is_relative_to (#3032).
- One audit row per write, delete and mkdir; tier refusals audited.
- Share holders' writes on an agent still on a pre-change image are refused
  until it restarts; agent keys pass and the row records it.
- Files tab reads the per-node flags instead of its own path lists; the
  Access tab states what a share grants.

Mutation-checked: reverting the tier gate fails 33 tests, the link-verdict
check 4, the containment fix 1; the old FilesPanel fails 7 of 8.

Fixes #3032

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…y source (ent#820)

The write deny lists moved from services/agent_service/files.py into
services/file_path_policy.py; the guard grepped files.py for the literal
paths and failed in CI. It now checks the policy's write verdicts and
still pins the guardrails-baseline.json entries.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…t#875)

A person an agent is merely shared with sat at the accessor tier for
execution reads and for the verbs that steer the agent: they listed,
opened and streamed every other person's turns, read the activity
timeline and the container log, terminated the owner's running turn,
changed the model, forced a git pull and minted or read webhook URLs.

Operator rulings (2026-10-15): one flat share grant, no schema change.
Share holders see only executions they started (lists, detail, live
stream, queued-prompt routes, fleet list narrowed in SQL before the
limit). Terminate needs owner/admin or the execution's own requester.
Webhook management is the schedule's creator or the agent owner. Model
change and git force-pull are owner tier; start/stop stay open.

- dependencies: viewer tier as an allowlist (owner / share / key), row
  scope helpers, assert_not_share_viewer (403 owner_tier_path)
- db: keyword-only source_user_id / own_rows_user_id with an exemption
  list for agents the caller provably owns; NULL never matches
- routers: schedules, chat, agents, git, executions, activities,
  fan_out, agent_config (capacity prompt previews)
- frontend: webhook controls hidden when can_manage_webhook is false;
  own-rows note on the Tasks panel
- scripts/management/list_webhook_rotation_candidates.py: read-only list
  of webhook schedules on currently shared agents, for rotation

Tests: 241 new/updated unit tests, route census + AST guard over the
read-helper call sites, real-SQLite narrowing proof, 18/18 mutations red.

Related to Abilityai/trinity-enterprise#875

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Review of #3776 found paths the policy still left at the share tier:

- the Python user-site (`.local/lib`, where the base image runs
  `pip install --user`): a `.pth` or `usercustomize.py` there runs as the
  agent at the next `python3`. Added to RUNS_CODE_DIRS.
- shell-init siblings missing from RUNS_CODE_FILES: .zshenv, .zprofile,
  .zlogin, .zlogout, .bash_login, .bash_logout.
- five fixed-name credential files readable at the share tier (the set
  predates this PR; the policy is now its single home): .netrc,
  .git-credentials, .docker/config.json, .config/gh/hosts.yml,
  .gemini/oauth_creds.json. Now owner tier to read.

Both policy copies stay byte-identical. Census and read-tier tests gain a
row per path; with the old policy 35 of them go red.

Left for a ruling: nested instruction files, bare repos not named *.git.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
@dolho
dolho force-pushed the feature/ent926-route-census branch from 99b73af to 7cdfa14 Compare October 16, 2026 13:50
@dolho dolho changed the title test(access): declare a tier for every agent route (ent#926) feat(access): route-tier census and per-turn file guard (ent#926) Oct 16, 2026
dolho and others added 5 commits October 16, 2026 17:47
…d-only owner-tier patterns (ent#820)

CI regression diff: test_files_protected_paths pins the read-only owner-tier
set as a literal so a new entry is a reviewed decision. The review-fix commit
added .netrc, .git-credentials, .docker/config.json, .config/gh/hosts.yml and
.gemini/oauth_creds.json without updating it. They are readable-only-at-owner
and writable like any file, same as .kube/config.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
- CodeQL: the rotation script prints token / signing-secret presence as
  constant text, so no credential-column value reaches the output.
- regression diff: test_the_fleet_route_passes_the_narrowing_to_the_query
  failed only under CI's test order. The fleet router binds viewer_own_rows at
  import, and another test reloads the helpers module, so the test patched one
  module copy and the route called another. The helper now patches the globals
  of every copy the code under test can reach.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Adds a static census over routers/: every route whose path names an agent
needs a row in tests/unit/agent_route_tiers.json, and a route with no row
fails CI, so a new route is owner tier unless its row says why not.

Rows are read against the code: an owner row must show an owner/admin gate
in the handler or its named helper, an accessor row must not, and a row
with no live route fails. The accessor rows tagged legacy are a shrink-only
ratchet (LEGACY_ACCESSOR_BASELINE).

No runtime change. Eight mutations (new undeclared route, dropped owner
dependency, accessor given an owner gate, renamed handler, baseline raised
or left high, stale path, emptied gate regex) each turn the census red.

Related to Abilityai/trinity-enterprise#926

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…llows (ent#926)

An agent acts for whoever chats with it and holds full access to its own
files, so whatever the Files API refuses a person the agent is shared with
they could ask the agent to do. This closes the file-tool half.

- hooks/turn-tier-guard.py (PreToolUse for Write/Edit/MultiEdit/NotebookEdit
  and Read/NotebookRead/Grep/Glob), registered in its own root-owned
  managed-settings.d drop-in so a bad entry cannot void the GUARD-002 file.
  It uses the Files API's own policy (third byte-identical copy, parity
  tested). A path the policy allows costs nothing; an owner-tier path asks
  the platform which tier the running turn acts for. Owner proceeds; share,
  unknown or no answer is refused.
- GET /api/agents/{name}/executions/{id}/turn-tier: the agent's own key
  only, answered from the execution row by turn_tier_service. A user decides;
  with no user, only turns the owner configured (schedule, reminder, loop,
  webhook, event) count as owner. Public links, channels, paid calls, voice
  and other agents get the restricted tier.
- The turn is identified by TRINITY_EXECUTION_ID in claude's launch
  environment (/proc/<parent>/environ), as the skill-gate already does.
- file_path_policy: on-disk conversation history (.claude/projects/*.jsonl,
  history.jsonl, .codex/sessions, .gemini chats) is owner-tier read, so the
  API and the hook agree. Project memory stays readable.

Not covered, and said so in security.md: a shell command reaches files the
file tools never see; Gemini/Codex runtimes run no Claude Code hooks; reads
of execution history through the agent's MCP key are not yet narrowed.

Tests: 93 new; eleven mutations (policy skipped, fail-open on unreachable and
on share/unknown, links not followed, no-run-id denied, shared person
promoted, missing user trusted, public trigger as owner, any key accepted,
policy copy drift) each turn them red. Stacked on the ent#820 policy.

Related to Abilityai/trinity-enterprise#926

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
regression diff and lint on the draft:
- test_2996 route census: classify the turn-tier route (agent-callable, the
  agent's own key and a path self-match).
- test_ent629 reach: turn_tier_service judges the person who started a turn,
  never the calling key; exempt with the reason.
- test_files_protected_paths: the four conversation-history patterns are
  reviewed read-only owner-tier patterns (pinned literal).
- sys.modules lint: the hook test serves lib and file_path_policy through
  monkeypatch for the module's lifetime instead of a bare pop.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
@dolho
dolho force-pushed the feature/ent926-route-census branch from 7cdfa14 to 44b0e67 Compare October 16, 2026 15:03
…s output path (ent#875)

CodeQL classifies by name: a dict key called has_secret was treated as secret
data even though only constant yes/no text reached the print. The fields are now
url_minted and hmac_on. Output unchanged apart from the labels (url / signing).

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
@github-actions

github-actions Bot commented Oct 18, 2026 •

Copy link
Copy Markdown

⚠️ Nightly unit-suite check skipped — merge conflict against dev.

Resolve by running git merge dev locally and pushing the result. The next nightly run will re-test once the conflict is gone.

dolho and others added 14 commits October 19, 2026 10:32
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…3295)

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…efuse it (ent#926)

- turn-tier-guard also covers Bash: in a platform turn only an owner-tier
  turn gets a shell (a command reaches files the file tools never see, and
  the container cannot host an OS sandbox).
- agent_server/services/turn_scope.py: Gemini CLI and Codex run no Claude Code
  hooks, so they refuse a share-tier turn before the CLI starts (403 naming
  the runtime; 503 when the tier cannot be checked). At spawn, so the push,
  sync chat and pull-mode paths all pass it; the backend still decides.
- The shared path policy classes three agent-owned paths outside the home that
  matched nothing: /app (runs code: on PYTHONPATH, startup.sh), the
  subscription token under /var/lib/trinity (credential), /logs (system).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…n it is (ent#926)

ent#875 left an agent key at the full view of executions. An agent acts for
whoever chats with it, so the auth dependency now records an agent key's
X-Trinity-Execution-Id (#2392) privately on the principal, and the viewer
helpers rate that turn with turn_tier_service: an owner-tier turn or no
platform turn keeps the full view; a share-tier turn reads that person's
rows (none with no platform user); an id that is not one of the key's
agent's executions reads none.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…' into feature/ent926-route-census

# Conflicts:
#	tests/registry.json
…nsus (ent#926)

With #3777 and #3776 merged in, 36 of the 152 legacy accessor rows already
have a ruling: the five shared-state reads and the model write are owner
tier, the execution reads narrow to the caller's own rows, the five webhook
routes are creator-or-owner, start/stop/model read stay open by ruling, and
the six Files routes are tiered per verb and path. Each row now carries that
reason; the legacy baseline drops to 116. The ent#875 census rules the new
turn-tier route (the agent's own key only).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…r through (ent#926)

Rulings 2026-10-19, verified live on a local instance:
- Visitors (channels, public links) vs shared people: a per-agent guardrail
  visitor_shell (default on; Settings > Guardrails, live, no restart) decides
  the shell and Gemini/Codex for visitors. Shared people never get them.
- The MCP server sends the turn on every backend call (AsyncLocalStorage in
  withAudit); per-method forwarding failed open.
- Standing work fence: an agent key acting for a shared person (or a limited
  visitor) cannot create or change schedules, loops, reminders, events or
  event subscriptions. emit-event is owner tier.
- A schedule's runs are rated by its creator, so a shared person's schedule
  is not the owner's. self_task is not an owner-configured trigger.
- Turns the agent dispatches carry the person their parent was for (one
  helper at every stamping site, AST-guarded).
- A runtime refusal is a typed 403 (turn_refused) end to end, not a
  retryable 503; refusal texts name visitors and shared people correctly.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…uard (ent#926)

turn_tier_service rates the person who started a turn, read from the
execution row, not the requesting principal, so no agent-key reach applies.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…uard (ent#926)

turn_tier_service rates the person who started a turn, read from the
execution row, not the requesting principal, so no agent-key reach applies.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… are told a refusal plainly (ent#926)

Found by the live run on 2026-10-19:
- fanOut (and four other client methods) called fetch directly, bypassing the
  turn header, so a fan-out from a shared person's turn stamped the owner.
  All backend calls now go through turnFetch; a guard fails on a bare fetch(
  in client.ts.
- A public-link visitor refused by the runtime (visitor switch off) was told
  'Failed to process your request. Please try again.' on both the sync path
  and the async status poll the public page uses. It now gets a plain,
  non-retryable line that names no internal setting (403 on the sync path).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@dolho dolho changed the title feat(access): route-tier census and per-turn file guard (ent#926) feat(access): hold an agent's turn to what its person may do, and declare every agent route's tier (ent#926) Oct 19, 2026
dolho and others added 2 commits October 28, 2026 14:44
Resolve tests/registry.json as a union of both sides' entries.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
trinity-enterprise#695 landed /api/systems/apply on dev after this census:
creator-gated, names no agent, and creates/updates members exactly as
/api/systems/deploy does, so it joins OWNER_TIER_UNNAMED beside /deploy.
test_ent926_agent_route_census flagged it unclassified after the merge.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@dolho

dolho commented Oct 28, 2026

Copy link
Copy Markdown
Contributor Author

Merged dev in again (2e23d1a21). The only text conflict was tests/registry.json, resolved as a union. I checked it per entry: no dev entry is lost, and each entry keeps the version from the side that changed it.

After the merge, test_ent926_agent_route_census::test_every_creator_gated_route_that_names_no_agent_is_fenced_or_declared_open went red on one route that dev added after this census: POST /api/systems/apply (trinity-enterprise#695). It is creator-gated, names no agent, and creates or updates members exactly as /api/systems/deploy does. So 8aecdf0e7 adds it to OWNER_TIER_UNNAMED beside /deploy, rather than to the open list.

Tests on the merged head:

  • The 27 test files this PR touches, plus the census and turn-tier files: everything passes except test_ent872_fleet_reach::test_the_operator_queue_reads_the_same_reach. That one failed once in the long combined run, passes on its own, and behaves the same way on fix(access): scope per-user views of agent activity (ent#875) #3777.
  • test_ent695_apply_route, test_agent_creation_role_gates and test_ent126_route_order all pass.

#3777 merged into dev (91a2cef). Its branch was rewritten before
landing, so its commits arrive again under new ids; the content equals
the 6424f75 tip this branch had already absorbed (b76b99a is the
same review-blocker fix). Every conflicted file therefore resolves to
this branch's version, which is #3777's final text plus #3778's own
extensions:

- turn_tier_service.py, executions.py, dependencies.py: #3778 extends
  #3777's code (Workspace channel/email columns, the standing-work
  fence for every agent key); no #3777 line is dropped.
- test_ent629/ent872 reach guards: the exemption follows the role check
  into _user_tier, where #3778 moved it.
- auth.md: #3778's new section 2.12 and the renumbered 2.13 self-gate.
- models.py: drop the duplicate _acting_execution_id PrivateAttr the
  auto-merge added (#3778 keeps it beside _acting_turn_cache).
- registry.json: dev's file plus this PR's six ent926 entries; drop the
  duplicated ent875 entries the auto-merge produced.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@dolho

dolho commented Oct 28, 2026

Copy link
Copy Markdown
Contributor Author

Merged current origin/dev (with #3777 at 91a2cef) into this branch: e7460a0. Pushed as a fast-forward, with no force-push and no rebase.

Why the conflicts happened. #3777's branch was rewritten before it landed, so its commits came back under new ids. For example, b76b99a is the same "close the three #3777 review blockers" fix as 6424f75, which this branch had already merged. For each conflicted file, dev's version is identical to #3777's final tip (4e6691c). When I re-ran the 3-way merge with the #3777 version this branch had already absorbed (6424f75) as the base, every code and test file resolved cleanly to this branch's version. So no #3777 review fix is lost and nothing actually contradicts.

Per file:

  • src/backend/services/turn_tier_service.py (add/add): kept this branch's version. Dev's copy is byte-identical to the 6424f75 version this PR extends with turn_actor, the Workspace channel/email rating and _user_tier.
  • src/backend/db/schedules/executions.py: kept this branch's get_running_turn_attribution. It is fix(access): scope per-user views of agent activity (ent#875) #3777's query plus source_channel and source_user_email, which the Workspace turn rating needs.
  • src/backend/dependencies.py: kept record_acting_turn (fix(access): scope per-user views of agent activity (ent#875) #3777) and added _enforce_standing_work_fence for every agent key (feat(access): hold an agent's turn to what its person may do, and declare every agent route's tier (ent#926) #3778 review fix).
  • tests/unit/test_ent629_agent_reach.py, tests/unit/test_ent872_fleet_reach.py: the turn-tier exemption is keyed on _user_tier, the function where feat(access): hold an agent's turn to what its person may do, and declare every agent route's tier (ent#926) #3778 moved the role comparison. The reason text is unchanged.
  • tests/unit/test_ent875_exec_control_routes.py, tests/unit/test_ent926_agent_key_turn_scope.py (add/add): kept this branch's version. Dev's copies match the 6424f75 base, so this PR's additions apply on top.
  • docs/memory/requirements/auth.md: kept this branch's version. It is dev's text plus the new §2.12 "A Turn Is Held to What Its Person May Do" and the self-gate renumbered to §2.13, with its two cross-references updated to match.
  • src/backend/models.py (auto-merged, fixed by hand): the auto-merge added a second _acting_execution_id PrivateAttr after _fleet_reach. I removed it and kept the one this branch places beside _acting_turn_cache.
  • tests/registry.json (auto-merged, fixed by hand): the auto-merge duplicated the four ent875 entries. I rebuilt the file from dev's version plus this PR's six ent926 entries. git diff origin/dev -- tests/registry.json now shows 62 added lines and nothing removed.

Migrations: this PR has none. Dev's Alembic track has a single head, 0104_agent_sync_state_git_stats_at, so nothing needed rechaining.

Checks:

  • python3 scripts/ci/check_alembic_heads.py src/backend/migrations/versions: 105 revisions, 1 head, PASS.
  • PR_BASE_SHA=origin/dev PR_HEAD_SHA=HEAD python3 scripts/ci/check_alembic_parity.py: pass, with no net-new MIGRATIONS entry.
  • pytest -p no:randomly passed: 2657 passed, 0 failed. It covered:
    • every tests/unit/*.py this PR changes;
    • every conflicted test file;
    • test_186_enumeration_uniformity.py, test_1310_auth_wiring.py, test_293_admin_gate_rejects_agent_keys.py;
    • the schema-parity trio (test_schema_parity.py, test_alembic_parity_guard.py, test_sqlite_upgrade_parity_guard.py).
  • git merge-tree --write-tree origin/dev HEAD: clean.

@dolho
dolho requested a review from obasilakis October 28, 2026 13:18

@obasilakis obasilakis left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Re-reviewed at e7460a0. Approving; CI is still running, so this assumes it lands green.

Status of my last review

  • Medium (a restricted turn reaching a create/deploy route that names no agent): fixed. route_tier_service.py:50-58 adds OWNER_TIER_UNNAMED (/api/agents, /api/agents/deploy-local, /api/systems/deploy, /api/systems/apply, /api/systems/{system_name}/restart), refused at :117 before the open branch at :119. test_ent926_agent_key_turn_scope.py covers shared turn refused, owner/scheduled/manual/no-header allowed, GET /api/agents open; test_ent926_agent_route_census.py:347 fails an unclassified creator-gated unnamed route or a stale entry.
  • m2–m6 and the owner-concurrency note: tracked in trinity-enterprise#978.
  • An AST scan of every mutating route with no agent in the path finds no create/deploy route left unfenced.
  • Merge e7460a0 (conflicts with #3777) checked with --remerge-diff: only conflict markers and duplicate definitions removed; each shared helper is defined once at head.

Low, optional

  • test_ent926_agent_route_census.py:332: the scan recognises require_role("creator") only as a handler argument. A create route gated through decorator dependencies=[...] would go unclassified, and unnamed routes default to open (route_tier_service.py:119-120). Nothing is affected today. Scanning decorator.keywords["dependencies"] too would close it.

dolho and others added 2 commits October 30, 2026 10:44
…urn tier

After merging dev:
- test_3408_{failed_handoff,gemini_process_ownership,gemini_reader_exit}
  landed after this branch stubbed the tier check in the other Gemini
  files; with no platform to ask, require_owner_tier fails closed (503)
  and the 13 regression-diff failures were that. Same owner-turn fixture.
- #4073 added template.yaml to file_path_policy.INSTRUCTION_FILES; the
  hook's byte-identical copy is refreshed.
- #4073's five metric-declaration routes take OwnedAgentByName; declared
  owner tier in agent_route_tiers.json.
- The creator-gated unnamed-route scan now also reads the decorator's
  dependencies=[...] (review, low).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@dolho

dolho commented Oct 30, 2026

Copy link
Copy Markdown
Contributor Author

Pushed e3e537d7b (on top of a fresh dev merge, d68ff0c34).

The red regression diff was this branch, not a flake. test_3408_{failed_handoff,gemini_process_ownership,gemini_reader_exit} landed on dev after this branch stubbed the tier check in the other Gemini test files. With no platform to ask, require_owner_tier fails closed (503), so the teardown paths those tests drive never ran. Fix: the same owner-turn fixture as test_2971/test_2433.

Also from the merge:

Review low (@obasilakis): _creator_gated_unnamed_routes now reads the decorator's dependencies=[...] too, not only handler args. No new route is classified because of it today.

Ran at head: the ent926 suites, the 3408 trio, #4064: 1323 passed. The registry conflict was resolved as a union. The rest of the unit suite is left to CI.

# Conflicts:
#	src/backend/services/chat_execution_service.py
#	src/backend/services/execution_envelope.py
#	src/backend/services/task_execution_service.py
@vybe

vybe commented Nov 1, 2026

Copy link
Copy Markdown
Contributor

merge-train (2026-11-01): ejected from this train — rides the next one once addressed. Validated at head 5123afa, lane C.

Needs intent, not mechanics:

  1. A voice-dispatched RUN_TASK on Gemini/Codex runs unrated at owner tier (gemini_voice.py:1488 builds the turn without a scope; turn_scope.py:94 only rates the Claude path), while an owner's own Claude voice tasks lose the shell. Either every runtime takes the scope or none does; today the two disagree in opposite directions.
  2. Contradicts fix(auth): agent keys act only as their own agent on self-gated routes; the agent-callable uses are listed #4040 (ent#719 pr1, which lands first on this train) on emit-event: fix(auth): agent keys act only as their own agent on self-gated routes; the agent-callable uses are listed #4040 narrows agent keys to their own agent, this PR widens the emit path. Also conflicts in src/ with it (event_subscriptions.py) and with fix(auth): an agent's settings, identity, git, credentials and keys take a person; their binds a signed-in session #4042 (agent_brain_orb.py). Re-merge dev once the ent#719 stack is on it and re-decide the emit path against fix(auth): agent keys act only as their own agent on self-gated routes; the agent-callable uses are listed #4040's rule.
  3. AC4 deviates by design (an owner's concurrent turn or schedule run is 403'd while any shared person's turn is running) — the body still says hold until the 1.0 cut, and there is no closing keyword.

Two operator questions carried for the ruling (the operator answers these on ent#926, not here):

@vybe vybe added the status-needs-fix PR has an unaddressed review/validation finding; cleared by the author's next push (#2815) label Nov 1, 2026
dolho and others added 2 commits November 2, 2026 17:32
…ensus

# Conflicts:
#	docs/memory/requirements/auth.md
#	src/backend/routers/agent_brain_orb.py
#	src/backend/routers/event_subscriptions.py
…v's gates

After merging dev:
- agent_route_tiers.json: the eight retrieval-memory routes (ent#1022).
  The six OwnedAgentByName routes are owner. The status read is accessor
  (status only; agent keys see only their own). The agent search door is self.
  list_agent_skill_gates now reads `via` assert_skill_gates_readable, which is
  where dev moved its gate.
- test_3180 census: turn-tier validates the agent's own key itself, so it is
  listed with OWN_VALIDATOR and driven in the sweep's door calls.
- test_ent926_turn_tier: an agent key emits only as its own agent since
  trinity-enterprise#719, so the reach test uses the agent's own key.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@dolho

dolho commented Nov 2, 2026

Copy link
Copy Markdown
Contributor Author

Merged dev into this branch to clear the conflict (9f82bb32b), then fixed what the merge broke (692e2abdd).

Conflicts

  • agent_brain_orb.py and event_subscriptions.py imports: kept both sides.
  • emit_event_for_agent: dev's assert_own_agent stays the first statement, followed by _assert_may_emit_for_agent, then this PR's owner tier for people. The comment about an agent key emitting for a sibling was dropped, because since trinity-enterprise#719 an agent key emits only as itself. The docstring now says a person may emit for an agent they own.
  • requirements/auth.md: kept dev's rewritten assert_own_agent row, renumbered to §2.13 to match this PR's new §2.12.

Merge fallout

Tests: locally, the ent#926, #3180, #2996, #837, #937, ent#719, #1310, #186, #293, ent#875, #2973 and brain-orb suites all pass: 4,558 tests, plus 705 on the re-run of the ent#926 census. One Alembic head.

@github-actions github-actions Bot removed the status-needs-fix PR has an unaddressed review/validation finding; cleared by the author's next push (#2815) label Nov 2, 2026
…owner's turn (ent#926)

test_ent1003_gemini_wall_clock and test_ent986_gemini_usage_statistics came in
with the dev merge and drive the Gemini runtime with an execution id. This
branch's tier check asks the platform whose turn it is, and with no platform
to ask it fails closed (503). These tests are about the CLI, not the turn, so
they take the same owner-turn stub the branch's other Gemini tests use
(turn_scope.fetch_turn).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@dolho
dolho merged commit 64033cb into dev Nov 3, 2026
30 checks passed
vybe pushed a commit that referenced this pull request Nov 3, 2026
…sus (#4206)

#2984 added GET /api/agents/{agent_name}/seat-autonomy and #3778 added the
census that requires every agent-naming route to declare a tier. They merged
one after the other without seeing each other, so dev fails
test_ent926_agent_route_census::test_every_agent_route_is_declared.

Declared accessor, with a reason, matching its sibling decision routes'
gates (assert_own_agent, then assert_agent_access).

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
vybe added a commit that referenced this pull request Nov 17, 2026
…4622)

* feat(agents): a tag in github:owner/repo@ref pins the clone; purge_agent_now removes an agent at once, audit first (#4183)

* feat(subscriptions): admission policy seam for a per-agent usage share (A)

One edition-agnostic point where a registered policy may stop an agent's
turn before it is admitted, on every door. With no policy registered
nothing changes. The first intended policy is an agent's share of its
subscription's usage, registered through this seam by an optional module.

- services/admission_policy.py: register / enforce / notice. A policy
  refuses with chat_signals.AdmissionRefused (code, plain-words message,
  frees_at); any other policy error is logged and the turn is admitted.
- /chat (admit_chat_request) and /task (dispatch_parallel_task) call it
  after the skill gate, before the idempotency claim: a refused request
  claims, uploads, writes and acquires nothing.
- execute_task step 1c covers every other door: the row closes SKIPPED
  with the message (not FAILED: no schedule failure count, no breaker),
  error_code usage_share, and the refusal on the result.
- routers/chat.py maps it to 429 + X-Trinity-Error-Code + Retry-After;
  a sync /task refused inside execute_task releases its claim and gets
  the same answer; the Workspace shows the message as category `limit`,
  not retryable.
- ExecutionContext.usage_notice renders the policy's notice as a
  `- **Usage**:` line via compose_system_prompt's auto-fill.

Tests: tests/unit/test_ent1059_admission_policy_seam.py (18). Each seam
call, the SKIPPED write, the router mapping, the Workspace branch, the
notice fill and the fail-open re-raise were removed one at a time and a
named test went red each time. test_2638's pre-dispatch allowlist gains
`policy_denied` with its reason.

Refs abilityai/trinity-enterprise#1059

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* feat(subscriptions): usage-share refusal survives the task backlog drain (B)

A share that ran out while a sync /task waited in the backlog: the drain
closes the row SKIPPED, and a long-poll that reads the terminal back from
the row (another worker drained it, or the wait timed out) rebuilt a
success-shaped result from a SKIPPED row. It now answers the seam's own
refusal.

- execute_task's policy gate stores the row error as `[usage_share] text`
  (the #3114 pull-sink shape), so a reconstruct can tell a share stop
  from any other SKIPPED; the in-memory result keeps the plain message.
- _dispatch_sync_backlog's reconstruct branch rebuilds AdmissionRefused
  from a SKIPPED row carrying that code; _map_task_failure then releases
  the claim and the router answers 429 usage_share (no Retry-After: the
  row has no frees_at, the message says when).

Tests: test_ent1059_admission_policy_seam.py +2 (20). The reconstruct
removed -> the drain test red; the prefix removed -> the SKIPPED-row test
red; each restored from a copy.

Refs abilityai/trinity-enterprise#1059

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

* feat(subscriptions): usage share card and breakdown column (C)

An agent's share of its subscription's usage, as its owner reads and
changes it. Both surfaces render only when the instance is entitled to
`usage_share`; an unentitled build calls nothing.

- Settings > Guardrails gains a "Usage share" section (UsageShareCard):
  the rolling 7-day and 5-hour windows as used / budget with a meter, the
  warning line drawn at warn_pct, Warning / Stopped badges in words, when a
  stopped window frees, whether a stop ask is pending, the line the agent
  reads on its next turn, and where each number comes from. The owner
  changes it inline through the PUT; an empty field is the platform
  default and all three empty clears the override. Bounds are named
  client-side before a request; a refused PUT keeps the form open with an
  InlineError. Applies from the next turn, no restart.
- Settings > Subscriptions per-agent breakdown gains "Share used (7d · 5h)":
  each agent's own percent, warning/danger tone from the share's own
  threshold; a share the reader cannot see renders a dash, never 0%.
- stores/agents.js: getUsageShare / setUsageShare (single API client).
  stores/subscriptions.js: fetchAgentShares, per-agent isolated.

Tests: src/frontend/tests/unit/usageShareCard.spec.js (22, mounted under
jsdom). Mutations each turned it red and were restored from a copy:
breakdown ungated (1 red), at_budget ignored (2), client validation
bypassed (5), Guardrails section ungated (1). Raw-colour ratchet: no
existing file grew; the new card carries 0 raw non-gray.

Refs abilityai/trinity-enterprise#1059

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(execution): every door answers an admission-policy refusal; a refused loop iteration is skipped, not failed (abilityai/trinity-enterprise#1059 fix)

Review C2, H1, the test_3274 sweep red and I5 on the pre-admission policy
seam.

C2: a turn an admission policy stopped inside `execute_task` is a SKIPPED
row `[usage_share] <sentence>`, and `_admission_policy_gate` returned it as
a result rather than raising — so every `dispatch_and_await_terminal` caller
that branched on failed/cancelled only answered success-shaped with an
empty reply. On the paid door that attached the execution to the claim,
logged verify-ok and SETTLED a charge for a turn that never ran.

- `chat_signals.admission_refusal_of(result, agent_name)` is the one rule:
  the in-memory attribute when the gate ran in this process, else a rebuild
  from the row's `skipped` + code shape. `result_from_execution_row`
  attaches it, so a turn refused on another worker (the backlog drain, a
  long-poll reading the row back) carries the same refusal; the separate
  rebuild in `_dispatch_sync_backlog` is that call now.
- One branch per door, before failed/cancelled: paid releases the claim and
  neither attaches nor settles (no charge; `payment.reason` is the
  sentence); MCP inline chat answers 429 with the sentence and
  `X-Trinity-Error-Code`; a channel sends the sentence and returns before
  step 12 so no empty assistant turn is persisted; the public link answers
  429 (sync) or stores the sentence as the reply (async); `portal_chat`
  reads status + code as well as the attribute; a room posts the policy's
  sentence and keeps its resume handle and cursor; A2A maps `skipped` to
  `rejected` with the message on the inline path as `_ROW_STATE` already
  did for the row; a refused validation turn is an ERROR verdict. The
  scheduler (`internal.py`), fan-out and operator-resume doors already
  record the skipped row with its message and need no branch; voice and
  the brain orb call `execute_task` directly and close SKIPPED.

H1: `loop_service._close_run` counted a policy-refused SKIPPED as a failed
run (only a gate-held SKIPPED was exempt), so `failed_runs` rose,
`on_failure` applied, and an at-budget loop kept ticking to `max_runs`. It
now reads the refusal off the row (`_split_error_code` + the predicate),
closes the run `skipped`, counts nothing, and stops the loop with the code
as `stop_reason`.

Sweep: `usage_notice` joins the platform-controlled list in
`test_3274_gate_context` — the line is composed by a policy from platform
numbers and never reads requester text. I5: `fetchAgentShares` runs at most
six reads in flight.

Docs: `requirements/security.md` §20.5d and `architecture/execution.md`
state the result-not-worker rule and each door's branch.

Red before the fix, with the door code reverted and the predicate kept:
all ten door tests in `test_ent1059_refusal_doors.py` (the paid one read
`settled`, the Workspace one DID NOT RAISE, A2A read `failed`) and both
loop variants beside the ent#751 hold test (`('failed', 'error')`);
restored byte-identical.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

* fix(completion-report): a refused turn's report shows the policy's sentence, not its [usage_share] code (abilityai/trinity-enterprise#1059 re-review N1)

An admission-policy refusal closes the row SKIPPED with `[usage_share] <sentence>`,
and the completion report (Workspace, Slack, Telegram) posted that text as-is.
The report now strips a known `[code]` prefix from a non-success terminal before
sanitising; an unknown `[...]` and any success text stay as the agent wrote them.
`_split_error_code` moves to the cycle-free `execution_envelope` leaf as
`split_error_code` so the report uses the same parser; the old names stay
importable from `task_execution_service`.

Also adds an executing spec for `fetchAgentShares`' six-lane cap (re-review N3):
a deferred `getUsageShare` over 15 names, high-water mark 6, every name written,
a failed read as `{ error: true }`. Mutation-checked: with the cap at 100 it
fails (expected 15 to be 6).

Red before the fix: test_the_workspace_report_shows_the_sentence_without_the_code,
test_the_channel_report_shows_the_sentence_without_the_code.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(agents): a dotted tag in github:owner/repo@ref pins the clone

The ref after `@` was checked against [A-Za-z0-9_-/] only, so a release tag
such as v1.2.0 failed the check and was dropped without a word: the agent was
created on main while the caller believed it pinned. The ref is now taken
whole when it is a valid branch or tag name (dots allowed, minus the git
ref-format shapes '..', a leading-dot component, a trailing '.' or '.lock'),
and an unusable ref is a named 400 instead of a silent fall-back to main.
The charset stays a shell-safety barrier for startup.sh's eval-built clone.

Red first: tests/unit/test_4165_github_ref_tag.py — the five dotted refs
returned url_branch=None and the ten unusable refs did not raise.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* feat(agents): purge_agent_now removes an agent at once, audit first (abilityai/trinity-enterprise#1055)

Agent delete is a soft delete (#834): rows, volumes and the MCP key stay for
the retention window. Removing an invited member's agent, or rolling back a
half-created one, must leave nothing but the audit trail, so the retention
sweep's hard purge becomes an immediate, edition-agnostic primitive with no
route of its own:

- the agent_lifecycle:purge audit row is written while the agent still
  exists and outlives it; if it cannot be written nothing is destroyed (503)
- a person caller must pass assert_admin (an agent key carrying an admin
  owner's role is refused); actor=None is the platform itself
- container, queued and in-flight work, Redis state, the #816 cascade
  (sharing, permissions both ways, MCP key) and the data volumes under both
  identities of a renamed agent (#1664) go; a base another agent claims
  stays; a Docker fault never stops the row purge
- system agents (403) and ephemeral agents (409) are refused; no row is a no-op

Red first: tests/unit/test_ent1055_purge_agent_now.py (module absent). Each
of five mutations turns it red: admin gate removed (2), fail-closed removed
(1), soft-delete-before-purge removed (7), base-claim guard removed (1),
audit moved after the container removal (2).

Docs: requirements/core-agent.md §1.4.1 and §4.2 ref pinning (trinity#4165);
architecture/agent-lifecycle.md pointer. Registry: both new test files.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(agents): a tag in github:owner/repo@ref survives the three ref-existence checks (trinity#4165, abilityai/trinity-enterprise#1055 review C1)

The parser accepted a dotted tag and stored it in `source_branch`, but every
check that consumed that value before the clone asked for a BRANCH of that
name: `provisioning.check_remote_branch_exists` ran `ls-remote --heads
refs/heads/<ref>`, `ref_probes._ref_arg` prefixed `refs/heads/`, and the
PAT path asked `GET /repos/{repo}/branches/{ref}`. So a tag passed the
parser and was refused with "Branch 'v1.2.0' not found" on every path; the
public doc claimed it pinned the clone.

Now all three accept a branch or a tag, the ref still a value after the URL
and never an option: the anonymous check asks for `refs/heads/<ref>
refs/tags/<ref>` with no `--heads`; the probes ask for both and take the
branch first (as `git clone -b` does), peeling an annotated tag to its
commit; the REST path answers a 404 on `/branches/{ref}` with
`/git/ref/tags/{ref}` before calling the ref missing. The refusal names
both: "Branch or tag '…' not found".

Red first: `tests/unit/test_4165_ref_existence_checks.py` drives the real
`check_remote_branch_exists`, `probe_anonymous_ref`,
`probe_tokened_repo_access` and `_validate_github_access` on all three
paths against a fake `ls-remote` that filters like git (so a check that
asks for heads only cannot see the tag) — 9 of 15 red before the change,
all green after. Three pins on the old branch-only contract moved with it
(`test_3725_*`: the argv now carries both refs and a same-named tag answers
when no branch does; `test_ent123_tokenless_clone`: the wording).

Docs: `requirements/core-agent.md` §4.2 names the three checks and the
precedence; the two flow rows that quoted `--heads` / the old wording
follow. Registry entry added.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

* fix(agents): a ref in github:owner/repo@ref may not begin with '-' (trinity#4165, abilityai/trinity-enterprise#1055 review I5)

`_GITHUB_REF_RE` admitted `-foo`. Safe today — every live site passes the
ref as a value (`git clone -b <ref>` takes the next token whatever it
starts with; the probes prefix `refs/…`) — but `startup.sh` also runs
`git checkout "${SOURCE_BRANCH}"`, where a leading `-` IS parsed as an
option, benign only because the clone already checked the ref out. One
shell edit away from not. The regex gains `(?!-)`; the 400 names it.

Red first: `TestRefRefused[-foo]` — DID NOT RAISE. Upgrade-safe: the parser
has one caller (creation); no stored `source_branch` is re-parsed on boot,
pull or reset, so existing agents are untouched.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

* fix(agents): an annotated tag in @ref resolves to its commit, not the tag object (trinity#4165)

`ref_probes` asked `ls-remote` for `refs/heads/<r>` and `refs/tags/<r>` and
took a peeled `refs/tags/<r>^{}` line when one was listed. git never listed
one: it shows a ref only when a pattern names it, and the peeled line's name
ends in `^{}`. So an annotated tag resolved to the TAG OBJECT, and the raw
template read was pinned to an id that is not the commit the clone checks
out. The probe now also asks for `refs/tags/<r>^{}` by name.

The unit fake listed `^{}` lines for a bare tag pattern, which is why the old
test passed. It now lists a ref only when a pattern names it, and a new file
runs the probes against a real bare repository over file:// (a lightweight
tag, an annotated tag, a branch and an annotated tag sharing a name).

Red first: test_4165_ref_probes_real_git.py — the annotated-tag and tokened
cases returned the tag object id; 4 others passed.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(agent-server): a tag-pinned agent's pull cycle is a named skip, not a failure every tick (trinity#4165)

`github:owner/repo@<tag>` clones with `git clone -b <tag>` in source mode,
which leaves the agent on a detached HEAD: pinned to that release, by design.
Pull sync is on for every `github:` agent, and `_run_pull_once` recorded
`failed: detached HEAD: nothing to pull onto` on every cycle, a failure
streak for an agent doing what it was asked to.

A detached HEAD now records `skipped` with the reason named. A detached HEAD
left by a rebase stopped part-way (its rebase-merge/rebase-apply directory
exists) stays `failed`, so a stuck tree still shows. The push cycle is
unchanged: it is off by default for a source-mode agent, and an owner who
turns it on for a pinned agent gets an honest failure.

Red first: test_4165_pull_cycle_pinned_tag.py::test_every_cycle_is_a_named_skip_not_a_failure
— 'failed' == 'skipped'. Real repos, the agent cloned -b v1.0.0 (annotated).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Trinity Agent (trinity) <trinity-agent@ability.ai>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: trinity-ability <309458136+trinity-ability@users.noreply.github.com>

* fix(credentials): keep the credential read inside the agent home and make the paste guard linear (#4191)

* fix(agent-server): the credential read stays inside the agent home (trinity-enterprise#1066)

CodeQL flagged the `os.lstat` this branch added to `GET /api/credentials/read`
(path depends on a user-provided value). The finding is real at the guard,
not at any current caller:

- the guard (`lstrip("/").lstrip(".")`, a `".." in` test on the stripped
  string, `str(resolved).startswith(str(home))`) let a single leading `../`
  through, and its prefix compare had no separator, so `../home2/.env`
  READ a file in a sibling directory whose name extends the home's;
- the new `lstat` ran before any containment check, so `../<anything>`
  came back in `missing` — an existence probe one level above the home.

Not reachable before this fix through the platform: the route sits behind
the agent's per-agent auth middleware (#1159) and every backend caller sends
fixed relative paths (`.env`, `.mcp.json`, `.credentials.enc`) or the agent's
own `/list` output.

Fixed at the root with one helper, `_contained_read_target`: refuse absolute
paths and any `..` component, then require both the normalised lexical path
and the link-resolved path to sit under the home with a trailing separator.
`lstat` runs on the contained lexical path (the absence proof); the read goes
through the contained resolved path, so a link is followed only to a target
that is itself inside.

Red first: tests/unit/test_ent1066_credential_read_containment.py — 5 of 13
red at the previous guard (sibling read, `../` oracle, absolute path, link to
a sibling file, linked parent directory), all green with the fix; the fix
reverted from a scratch copy turns the same 5 red.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(sanitizer): the paste guard's KEY=value walk is linear by construction (trinity-enterprise#1066)

CodeQL flagged the two regex calls in `secret_shape` (py/polynomial-redos,
"may run slow on many repetitions of '!'"). Origin: the patterns pre-exist
on dev — `_KV_LINE_RE` / `_KV_PAIR_RE` are the redaction's #1661/#3311 walk —
but these two alerts are introduced by this branch, which feeds them a
Workspace chat message (user-provided text) through the new guard.

Measured before the change, CPython was already linear on every shape tried
(`!`*n, `=`+`!`*n, `!`*n+`=`, `!=`*n, `k=!`*n, an unclosed quote, sensitive
keys with short values): 1M characters took 0.04-1.1 s, scaling ~10x per
10x. The lookbehind pins each key to the start of its run — the engine
detail CodeQL's model does not see (#1661 says the same). The route caps a
message at 100k characters.

The guard no longer rests on that detail. It finds each `=` with
`str.find`, walks back to the key start (never past the previous `=`, so
no character is visited twice), and checks the value with one fixed-width
match — an optional opening quote and eight non-space, non-quote
characters. No length cap, so nothing at the end of a long message is
dropped. One deliberate widening: the closing quote is no longer required,
because a truncated paste is still the secret. The redaction path is
unchanged.

Red first: tests/unit/test_ent1066_secret_guard_linear.py — red before the
change on the no-unbounded-regex pin and the unclosed-quote case (2 of 43);
the timing bound (1M < 3 s, 10x input < 30x time) and both-way detection
were green before and after. The change reverted from a scratch copy turns
the same 2 red. The 1661/2208/2398/3840 sanitizer suites and the paste-guard
suite stay green (608 passed).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Trinity Agent (trinity) <trinity-agent@ability.ai>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* chore(enterprise): bump the submodule pointer to main (448cdbc -> f347a02) (#4193)

Picks up abilityai/trinity-enterprise#1072 (member_agents module, ent#1055),
which composes the purge_agent_now primitive from #4183.

Co-authored-by: trinity-ability <309458136+trinity-ability@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* revert: an agent's share of a subscription (#4180) (#4199)

* Revert "feat(subscriptions): an agent's share of a subscription, enforced at every door (#4180)"

This reverts commit 9eedbbb6e0be8d5b73ccb5e8f0217a33a987b089.

* chore(enterprise): bump the gitlink to the ent#1070 revert (trinity-enterprise#1076)

Lands with the #4180 revert so dev never carries the usage_share module
without the admission_policy seam it imports.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: trinity-ability <309458136+trinity-ability@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* fix(sub-003): the agent's auth answer outranks a prompt-length mention, and the turn-limit anchor is pinned (#4123) (#4197)

* fix(workspace): keep a room's content inside the Inbox reading pane (#4160) (#4169)

Every agent bubble renders PortalMarkdown's sr-only live region, which is
position:absolute. The reading pane's scroll box was not positioned, so
those regions resolved against the page, escaped the overflow box, and a
long room stretched the document: the wheel scrolled the whole Workspace.
Make the scroll box `relative`, as the list column's already is.

Adds a mount spec pinning the containment structure for a long room, a
room with many messages, and a thread.

Fixes #4160

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* refactor(operator-queue): the listing engine moves to listing.py, and the poller core into PollerCoreMixin (#3785, PR 2 of 4) (#4037)

* fix(operator-queue): the #1677 helper delegates to the platform-alert seam

create_bounded_alert(_outcome) keep their signatures and outcome contract
and become a thin delegate of platform_alerts.observe, which already
composes the per-(agent, type) budget inside its one locked create. The
registered type names its kind; the key is read back from the item's id
and context by derive_legacy_subject and must resolve to that same kind,
else ALERT_FAILED_CREATE. The two kinds whose ids carry no subject
(gitignore, problem report) are keyed by the request id minus the kind's
prefix. observe's "updated" and "snoozed" pass through; "failed" maps to
ALERT_FAILED_CREATE; every refused or failed arm logs [#1677 budget].
ALERT_FAILED_COUNT stays importable and is no longer returned.

The budget episode reports through observe as kind alert_budget keyed by
the budgeted type: one alert_budget:<type> row per agent, re-read by a
later refusal outside the cooldown, ended by its 14-day lifetime. The
bucketed id and its own operator_queue_new broadcast are gone; observe
sends the thin operator_queue_sync trigger.

Ratchet: create_bounded_alert_outcome and _maybe_emit_alert_budget_episode
leave EMITTERS_NOT_YET_ON_SEAM; alert_budget gets a NO_CLEAR_SITE reason.
Census: both leave _ALLOWED_CALLERS (no direct create left in either).

#3254 re-pins (asserted outcomes unchanged):
- test_1677_operator_alert_budget: write-observation point. The harness
  stands in for the seam's create with the same budget arithmetic and
  runs what observe spawns. Id shape in
  test_deterministic_bucketed_id_with_reserved_prefix,
  test_cooldown_knob_zero_never_raises (the subject replaces the bucket)
  and test_under_cap_creates_item_and_notification (the seam's id). The
  episode raise now stays in its spawned task, so
  test_episode_helper_raise_swallowed_refusal_stands reads the delegate's
  WARNING marker; test_ws_broadcast_failure_swallowed checks the thin
  operator_queue_sync payload.
- test_3130_flood_guard, test_1632_operator_queue_caps: write-observation
  point. The flood alert still rides the delegate, so its writes, the
  typed budget read and the I1 fault injection move to the seam's create.
- test_3243_atomic_asks::test_a_cap_hold_never_fires_the_flood_alert:
  write-observation point. Watched at observe, since the oqs.db assertion
  alone would now pass whether or not the alert fired.

Known until the problem report moves in its own commit: nine
test_3521_problem_report_comment cases are red here. The delegate files
an event-kind row under the seam's id, which the old _attach_comment
(keyed on the pre-seam request id) cannot find.

New: tests/unit/test_3254_budgeted_alerts.py (registry parity, reach-back
imports, the delegate's outcome table over all five budgeted types, the
episode's seam properties, R1 for alert_budget).

Part of #3254

* fix(operator-queue): the flood alarm reports through the seam and ends with its episode

OperatorQueueSyncService._maybe_emit_flood_alert reports through
platform_alerts.observe as kind queue_flood keyed by the reason: one
pending row per (agent, reason), rewritten in place, budgeted by the
seam. #3130's outcome rule is kept on observe's strings: a failed write
releases the episode and is retried after the cooldown (I1); created,
updated, snoozed and refused consume it.

While an episode lasts it is re-read at most every 6 hours, so the kind's
14-day lifetime ends only a row nobody re-read, never a condition that
still holds. The refresh time per (agent, reason) is a dict the method
creates on the instance for itself; __init__ is unchanged. The flood's
own operator_queue_new broadcast (which carried the title) is gone;
observe sends the thin operator_queue_sync trigger.

The two _sync_agent statements where an episode ends
(_flood_episodes.discard for oversize_file and ingestion_cap) become
calls to the new _end_flood_episode, which ends the episode and, only
when one was running, ends its alert as the platform through
platform_alerts.clear. A cycle with no episode costs no read.

Ratchet: _maybe_emit_flood_alert leaves EMITTERS_NOT_YET_ON_SEAM;
queue_flood's clear site is _end_flood_episode.

No re-pin in this commit: the #3130 and #1632 harnesses were moved to
the seam's create in the previous commit and keep their results.

Tests: the flood through the real _sync_agent poll for 30 simulated days
with the real expiry sweep (one row, re-read every 6 hours, ended as the
platform when nothing is held), no read without an episode, two reasons
two subjects, I1 on the real seam, a person's ending standing through the
re-read and the clear, the thin trigger, R1 for queue_flood.

Part of #3254

* fix(execution): the effect guard and missing-skill alarms report through the seam

_on_unguarded_effect reports a refused effect through
platform_alerts.observe as kind effect_unguarded keyed
"<effect_type>:<reason>": one row per agent, effect and reason,
rewritten by each further refusal, ended by its lifetime (nothing
observes an agent starting to send ids). The send is refused whatever
the alarm's outcome.

_alert_skill_not_found reports through observe as kind skill_not_found
keyed by the truncated command: one row per agent and command, re-read by
each scheduled run. The paired notification is sent only on "created": a
repeat of a pending command is an "updated" reading and already had one;
refused, failed and snoozed wrote no new row. The pre-create dedup read
stays and now matches only a pending pre-#3246 row (no subject), which
keeps standing for its command until the upgrade sweep stamps it; a seam
row is re-read through observe, where a repeat is never refused at the
cap.

task_execution_service.py: only _alert_skill_not_found changes. Its
module-level utc_now_iso import is now unused and is left in place.

Ratchet: _on_unguarded_effect and _alert_skill_not_found leave
EMITTERS_NOT_YET_ON_SEAM; skill_not_found gets a NO_CLEAR_SITE reason
(only the dispatch path could see a command resolve, and there the
detector cannot tell "resolved" from "not a slash command").

#3254 re-pins (asserted outcomes unchanged):
- test_1677_operator_alert_emitters::
  test_skill_not_found_emitter_is_routed_not_allowlisted: the module must
  call platform_alerts.observe naming "skill_not_found" (AST), the seam
  that composes the budget, instead of mentioning create_bounded_alert.
- test_2392_effect_guard_fail_closed: write-observation point and id
  shape. The alarm is recorded at platform_alerts.observe instead of a
  sys.modules stub of operator_queue_service, which lacks the names the
  seam reaches back for. The id prefix is asserted on the kind; the
  at-budget case returns the seam's refused_at_budget.
- test_ent879_aauth_resource_route (the refused-call case): the dead
  create_bounded_alert patch moves to platform_alerts.observe.

Tests: the effect refusal's seam properties and lifetime; the missing
skill's one row per command, no notification on a repeat, a pre-seam row
still standing, no notification at the cap; R1 for both kinds; the other
shape, the clock 30 days ahead: a full skill_not_found budget frees
itself through the real expiry sweep and a new command files.

Part of #3254

* fix(git-sync): the gitignore sweep's alert is one row per change set

_emit_gitignore_untracked_alert reports through platform_alerts.observe
as kind gitignore_untracked, which stops being an event kind
(platform_alerts.py: event=True dropped). It is a must-act kind: no
lifetime, never ended by the platform.

It is keyed by the change itself: a 32-hex digest of the canonical,
direction-tagged set of paths the Push untracked and newly un-ignored
(sorted removed paths, the exact removed count, sorted un-ignored paths).
A repeat of the same set collapses onto one row; a different set, a
different total behind a capped list, or the same path in the other
direction is a different row. Keying per agent would let a later Push
overwrite an earlier Push's path list on a pending must-act alert, which
#2529 calls the surface that outlives the session. The legacy prefix
keeps its _no_subject parser, so pre-existing rows stay pending until a
person acts.

Ratchet: _emit_gitignore_untracked_alert leaves EMITTERS_NOT_YET_ON_SEAM.

#3254 re-pins (asserted outcomes unchanged):
- test_2529_gitignore_precedence: write-observation point and id shape.
  The alert is watched at platform_alerts.observe (the two
  create.await_count == 0 checks would otherwise have gone vacuous); the
  per-agent id prefix is asserted on the kind, with the 32-hex key.
- test_3246_platform_alerts_leaf::test_event_kind_has_no_subject:
  gitignore_untracked is a condition now, so it is asserted keyed and the
  problem report stays the event example.

Tests: a repeated set is one row that never expires; distinct sets and a
capped list with another total stay distinct; the direction tag holds;
a person's ending stands; the 30-day case now also shows the gitignore
row still pending.

Part of #3254

* fix(workspace): the problem report is one seam row per person, target and day

raise_problem_report reports through platform_alerts.observe as kind
workspace_problem_report, which stops being an event kind (event=True
dropped). It is keyed by the 32-hex digest _problem_report_id already
computes: one row per person per target per UTC day, the "never
duplicates" rule its docstring states. A person must act on it, so it
never expires and the platform never ends it. The workspace-problem- id
prefix is kept, so ent#866's owner-only rule and the machine-key "about a
person" exclusion still hold. It returns True when the row was created or
updated, False when snoozed, refused or failed.

#3521's words-only rule moves under the seam's lock. observe gains
keep_text (passed through database.create_platform_operator_queue_item to
create_platform_item and _touch): with keep_text and a pending row, the
reading only counts (last_seen_at, seen_count, expires_at) and the title,
question, priority and context stand. The bare click passes keep_text, a
reading with words does not. The click and the comment are two background
requests that meet on one row inside the per-agent lock, so either order
ends with the words. With no pending row the reading files as usual.

update_pending_problem_report (database.py, db/operator_queue.py) and the
nested _attach_comment are removed: one write path.

Ratchet: raise_problem_report leaves EMITTERS_NOT_YET_ON_SEAM, which
fixes the nine test_3521 cases the delegate commit left red.

#3254 re-pins (asserted outcomes unchanged):
- test_3521_problem_report_comment: the acknowledgement fixture goes
  through ask_service.answer (fixture shape; a raw status UPDATE carries no
  person ending for the snooze to read); the accessor test asserts the
  same type guard at the seam (a foreign row under the report's old id is
  never written); the lost-race test injects the bare click's whole
  reading, in its own thread, before the comment's locked create.
- test_ent499_problem_report: write-observation point. The fake moves from
  create_bounded_alert to platform_alerts.observe; the routing AST test
  asserts the call to observe.
- test_ent715_queue_person_fields: update_pending_problem_report leaves
  _NON_ROW_ACCESSORS (the accessor census forbids a stale entry).
- test_3246_platform_alerts_leaf::test_event_kind_has_no_subject: no
  built-in kind is an event any more, so the property is pinned on a kind
  registered for the test, and both former events are asserted keyed.
test_ent866_problem_report_owner_only is unchanged.

Tests: keep_text at the sink, both interleavings keeping the words, one
must-act row per person, target and day, the owner-only prefix, the
delegate and the emitter landing on one subject, a person's ending not
rewritten or reopened; the 30-day case now also shows the report still
pending; the five-type admission test asserts every kind keyed.

Part of #3254

* fix(operator-queue): decide a platform alert's snooze under the create's lock

observe checks the snooze before the create's lock. Two moments let a
person's ending slip past that check: an acknowledgement after observe
read the pending row but before the locked create (the locked find then
misses), and one between the locked find and _touch's compare-and-set
(the lock covers creates, not respond_to_item, so the CAS loses). Either
way the create fell through to an insert, and the alert a person had just
ended was back at once. Before this batch most of the moved writers had a
deterministic id that could not re-file after an ending, so the moves
would otherwise ship that regression.

create_platform_item gains snooze_since and a snoozed predicate. On the
insert path, after a missed find or a lost CAS and inside the same locked
transaction, it reads the newest person-ended row for the subject through
the transaction's own connection (the public lookup opens another one)
and returns {"outcome": "snoozed"} without inserting when the predicate
says so. observe passes the standard 7-day window, computed on its own
rather than from the pre-lock check's conditionally assigned cutoff, and
lambda ended: _snoozed_by(kind, ended, priority, context). The pre-lock
check stays as the fast path. A caller that passes neither keeps the old
behaviour.

The statement behind the public find_person_ended_by_subject becomes
_find_person_ended_by_subject(conn, ...), shared by both reads.

Tests: a person's acknowledgement through ask_service.answer, landed (a)
between observe's read and the locked create and (b) between the locked
find and the CAS, yields "snoozed" and no new row, for the problem
report, an inbox collision as registered on dev, and the flood alarm's
6-hourly re-read. All six cases were red before this commit (a fresh row
was created) and pass with it. On SQLite, BEGIN IMMEDIATE holds every
other writer, so moment (b) lands the ending first and hands both finds
the row as it stood just before it; two real workers are not exercised.

Review fixes carried here (the earlier commits are left as they are):
- OperatorQueueSyncService._maybe_emit_flood_alert: the cooldown is
  stamped before every reading, the 6-hourly re-read included. A failed
  re-read released the episode with the cooldown stamp still 6 hours
  old, so it was retried on the next 5-second cycle. New test: a failed
  re-read is retried no sooner than the cooldown (red before the fix).
- test_3521_problem_report_comment (the lost-race re-pin): asserts the
  bare click's own request returned True and left a pending, comment-free
  row before the comment's locked create ran, so the test cannot pass
  without the interleaving it claims.
- gitignore_sweep: states that past the probe's 2,000-line cap the key is
  the capped list plus the exact total (a full-set digest needs the
  in-container probe, outside this change).
- test_1677_operator_alert_emitters: the census's guidance for an
  agent-influenceable emitter now points at platform_alerts.observe, not
  the delegate the #3246 ratchet counts as off the seam.
- test_3254_budgeted_alerts: the frozen clock also covers
  db.operator_queue's ent#815 commit-lag check, which logged spurious
  errors against the real clock.

Part of #3254

* fix(operator-queue): a failed flood re-read keeps the episode, so its end still clears the alert

Fresh review of the #3254 budgeted slice (two findings):

- A failed 6-hourly re-read of a persisting flood released the episode.
  If the condition then cleared before the retry, `_end_flood_episode`
  found no episode and the pending alert waited out its 14-day lifetime,
  although the code had seen the condition clear (AC 2). A failed re-read
  now keeps the episode, whose alert an earlier reading filed, and the
  re-read is retried once the cooldown has passed. A failed first reading
  still releases the episode (#3130 I1). New test:
  test_a_failed_reread_keeps_the_episode_so_its_end_still_clears_the_alert
  (red with the old release branch put back).
- AC 8 for two kinds had no "a person's ending is never overwritten" test:
  `skill_not_found` (the next run is snoozed and sends no notification)
  and `alert_budget` (a later refusal re-reads the episode and the
  acknowledgement stands). Both added.

* test(operator-queue): the budgeted-alert suite freezes the clock its running methods read

In the full serial unit tier, test_1595_sync_health_signals evicts
`database` and every `db.*` module from sys.modules and restores only
some of them. Afterwards `database.db` and `ask_service.db` run on an
older `db.operator_queue` module than the one this suite's clock fixture
imports and patches, so the seam's create, the expiry sweep and the
answer's compare-and-set read the real clock: rows filed at the frozen
T0 read as long expired (five cases red, `AskConflict: expired`, a
second flood row). The fixture now also freezes `utc_now_iso` and
`datetime` in the globals of the live operations class (R6: patch what
the running method reads). Red before: test_1595 followed by this file,
7 failed; green after: 66 passed; the file alone: 55 passed.

* refactor(operator-queue): operator_queue_service becomes a package, content unchanged (#3785)

The 2,840-line module moves to services/operator_queue_service/__init__.py
byte for byte (100% similarity), the first step of the Invariant #1 split.
Every `from services.operator_queue_service import X` and
`from services import operator_queue_service` resolves as before.

Six path-keyed guards name the file by its path, so their key moves with it
(the PR #2487 precedent): _NOT_EMITTERS (test_3246), _ALLOWED_CALLERS
(test_1677), _ALLOWED (test_1920), _OQS (test_ent715), FLEET_LEVEL_ALLOWLIST
(test_ent467), and the source read in test_ent499. Only the key changes; the
rule, the reason text and every assertion are untouched.

Part of #3785

* test(operator-queue): pin the package's one-binding rules before the split (#3785)

tests/unit/test_3785_operator_queue_package.py pins what S2–S6 must keep
while they move code out of __init__, each property derived from the tree
so no later slice edits it: modules under 800 lines (__init__ exempt, with
a staleness check); the import surface src/backend and tests/ read still
resolves; no from-import between package modules; one top-level binding per
name; import-time reads defined first and the _MARKER_UNCHANGED default the
owner's object; a patch of a re-exported name lands where its readers read;
one logger identity; owner-exact qualified reads; no submodule named like a
package attribute; one rate_limiter; no method on two classes of the
composed service; external imports before the first submodule import,
and a failed first import, injected at that line, leaves no stale submodule.
The AST scans live in _oqs_package_scan.py (pure stdlib, never imports src).

The other shape, a fresh install: each consumer imported first in its own
interpreter on an empty database raises a bounded alert (created; only main
binds a WebSocket manager, and then exactly one thin operator_queue_sync
trigger goes out). Outcomes equal on the base.

Part of #3785

* refactor(operator-queue): bind the service's collaborators once, in deps.py (#3785)

db, _queue_db, principal_reach, get_breaker_redis, ask_service, AgentClient
and the audit sink _audit_sync move to services/operator_queue_service/deps.py,
bound once and read as deps.X everywhere in the package. They are no longer
attributes of the package, so a stale monkeypatch.setattr(oqs, "db", ...)
raises AttributeError instead of patching a copy nothing reads (Invariant
#1(a), the PR #2487 rule). rate_limiter stays on the package: ask_service
reads oqs.rate_limiter.check.

Tests that patched a collaborator on the package now patch oqs.deps.X; no
assertion, expected value, fixture or control flow changes. The package guard
gains G4 (deps binds exactly the collaborators, no other module imports one
from its source under any alias or holds one by identity, the package has
none) and refuses any reach of a package function's __globals__.

Part of #3785

* refactor(operator-queue): the queue listing engine moves to listing.py (#3785)

QueueListError, narrow_to_agent_key, the cursor and walk-snapshot helpers and
list_for_principal (trinity-enterprise#815) move whole from __init__ to
services/operator_queue_service/listing.py. The two package-owned prefix
tuples are read as _oq.X, collaborators as deps.X. The package re-exports
exactly the six names read from outside (the router's list_for_principal and
QueueListError; narrow_to_agent_key, encode_cursor, decode_cursor and
CURSOR_MAX_BYTES, which tests call or compare).

test_ent715's row-reader census re-keys list_for_principal to listing.py; no
assertion changes.

Part of #3785

* refactor(operator-queue): the poller core moves into PollerCoreMixin (#3785)

The instance state, start/stop, the Redis leader lease, _poll_loop,
_poll_cycle and _broadcast_sync move into PollerCoreMixin in
services/operator_queue_service/poller.py. OperatorQueueSyncService is now
composed in __init__ from the mixin and keeps its singleton there (the
db/schedules shape). Package-owned names (DEFAULT_POLL_INTERVAL, _LEADER_KEY,
_RESERVED_ID_PREFIXES, _websocket_manager) are read as _oq.X; the default
poll_interval still binds 5 at import.

Three path-keyed censuses re-key their poller entries to poller.py
(test_1920's lease, test_ent467's operator_queue_sync trigger, test_ent715's
_poll_cycle); no assertion changes.

Part of #3785

* fix(operator-queue): a new flood, budget or untracking episode files after an acknowledgement (#3254)

A person's ending of a `queue_flood`, `alert_budget` or `gitignore_untracked`
row snoozed every later reading of the subject for 7 days, because none of
the three kinds had a material key, so `_snoozed_by` was vacuously true. A
runaway agent's second flood, a second budget-full episode and a second
untracking Push of the same paths were all silent (review C1; dolho on
operator_queue_service.py:2800 and gitignore_sweep.py:335).

Each reading now carries an episode marker that the kind lists in its
`material_keys`:
- queue_flood: `episode_started_at`, stamped when the in-memory episode
  starts. Re-reads of one episode share it; an episode that starts after the
  condition cleared does not.
- alert_budget: `filled_by`, a digest of the ids of the pending rows filling
  the budget. Re-reads of the same full budget share it; once a slot frees
  and a new row fills it, it differs.
- gitignore_untracked: `push_id`, one per call. The emitter runs once per
  Push, so every untracking Push after an acknowledgement files; repeats
  still collapse onto a pending row.

So the snooze silences re-reads of the acknowledged episode for the snooze
window and a new episode files at once, which makes the
`_maybe_emit_flood_alert` docstring's #3130 edge-trigger claim true again.
The budget episode's read of the filling rows is classified in the ent#780
row-reader census as internal.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HWna79vVmnezQdhpaGRKbw

* fix(operator-queue): a flood re-read no longer stamps the per-agent cooldown; the oversize end is pinned (#3254)

Review follow-ups on the flood alarm:
- dolho on operator_queue_service.py:2778 (webmixgamer agreed): the 6-hourly
  re-read of one reason's episode stamped the per-agent
  `_flood_alert_cooldown`, so the other reason's first alert could wait up to
  one cooldown. The cooldown spaces episodes and a re-read is not one, so it
  is now stamped only before an episode's first reading. A failed re-read's
  retry spacing already comes from its refresh time.
- webmixgamer I3: the 6-hour re-read spacing is one constant,
  `OPERATOR_QUEUE_FLOOD_REREAD_SECONDS`, read by the gate and by the failed
  re-read's retry arithmetic, so the two cannot drift.
- webmixgamer I1: a `_sync_loop` case serves an oversized file and then a
  sane one and asserts the row ends as the platform. It goes red when
  `_sync_agent`'s `_end_flood_episode(agent_name, "oversize_file")` is
  replaced by a bare episode discard.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HWna79vVmnezQdhpaGRKbw

* docs(operator-queue): state the fail-open double-leader bound on the flood episode marker (#3254)

The episode marker is worker-local. With Redis down the leader lease fails
open, so each worker stamps its own episode_started_at and another worker's
re-read can file past a person's ending, at most one extra row per extra
worker per re-read window. Docstring only; no behaviour change (review L1).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HWna79vVmnezQdhpaGRKbw

* lint(operator-queue): __init__ drops its file-level F401 waiver (#3785)

Review I1 (dolho): the file-level `# ruff: noqa: F401` switched off
unused-import detection for all of `__init__`, so a dead import left behind
by a later slice or feature would go unflagged. At this PR's head
`__init__` still reads every name in its import block, so the waiver hid
nothing: `ruff check --select F401` gives 0 findings with it removed. It is
dropped, and the block's comment now says that a name `__init__` stops
reading keeps its line with a per-line waiver naming why (G13, or read as
`oqs.X` from outside). PR2, PR3 and PR4 add those per-line waivers where
their moves leave a name unread.

`deps.py` keeps its file-level waiver: it is a binding module whose
imports are its bindings, so there is no import there F401 could rightly
flag.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HWna79vVmnezQdhpaGRKbw

* lint(operator-queue): per-line F401 waivers for what __init__ stops reading (#3785)

Review I1 on #4036: PR1 dropped `__init__`'s file-level F401 waiver. This
slice's moves leave four of its imports unread in `__init__`; each now
carries its own `noqa: F401` with the reason, so F401 keeps checking every
other line:
- `asyncio`, `base64`, `uuid`: kept for G13, because `listing` and `poller`
  import them at module level.
- `from . import listing`: loads the submodule at its G6 point; the names
  callers read come from the re-export block.

`ruff check --select F services/operator_queue_service/` is clean.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HWna79vVmnezQdhpaGRKbw

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* refactor(frontend): useVoiceSession goes through the single API client (#2720) (#4185)

The last bare axios call in the composable, voice stop, built its own auth
header. It now uses the shared client, so it follows the interceptor's
token and 401 handling (Invariant #7). The WebSocket URL still takes the
store token, which is a different path.

Closes #2720

Co-authored-by: trinity-ability <309458136+trinity-ability@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* test(memory): stub last_error in the remaining memory_on fixtures (#4187)

The memory_on fixtures in test_ent1022_retrieval_memory and
test_ent1027_model_gateway stubbed is_loaded and availability but not
last_error, so an embedder error left by an earlier test on the same
worker made status() report memory unavailable. The 1028 fixture got
the same stub in #4145.

Fixes #4137

* docs(readme): link the Trinity 1.0 complete course (#4063)

* docs(readme): link the Trinity 1.0 complete course

Add the 39-chapter complete course video as the first entry under
"Watch more" so new visitors find the end-to-end walkthrough first.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* docs(readme): update Complete Course link to new YouTube id and chapter count

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* docs(readme): update Complete Course YouTube link to re-published video

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

---------

Co-authored-by: Trinity Agent (trinity) <trinity-agent@ability.ai>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

* chore(metrics): code-health dashboard 2026-11-03 @ 93658be60 (#4198)

Co-authored-by: Trinity Agent (trinity) <trinity-agent@ability.ai>
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>

* feat(auth): terms acceptance gate and invite door (trinity-enterprise#1054) (#4195)

OSS half of trinity-enterprise#1054: the terms-gate seam (services/terms_gate.py, consulted in resolve_platform_user_unfloored; @terms_route exemptions), EmailService.send_html on one HTML shell, the /terms SPA view + interceptor, and the entitlement-gated InvitesPanel.

Refs abilityai/trinity-enterprise#1054

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* feat(workspace): a Direct call — every spoken turn goes to the agent and you hear its own reply (trinity-enterprise#1046) (#4194)

* feat(workspace): a Direct call, where every spoken turn goes to the agent and you hear its own reply (trinity-enterprise#1046)

A second call mode beside Live, chosen from the Call control's menu (remembered per
browser; Live stays the default). No model sits in the conversation: the browser
endpoints an utterance, the backend transcribes it, and the browser sends the text
through the same POST .../chat/stream a typed message uses, carrying the call's vd_ id.
The bridge then speaks the agent's reply sentence by sentence.

- Harness-agnostic: every reply is spoken from the persisted row; a runtime whose
  stream carries messages (Claude Code, Codex) only starts sooner, and the unspoken
  tail is reconciled so nothing repeats. Gemini CLI agents work, with a line saying
  replies take about 10 s.
- Agent approvals by voice: an ask the turn raises is read aloud, and a whole-utterance
  yes/no after the prompt has played is posted by the browser on the person's own
  answer route. The socket never answers an ask and never starts a turn.
- Gates: socket owner only (no admin attach); a turn may carry only the caller's own
  live Direct call on this agent and thread (uniform 404); a spoken run must be the
  caller's own execution, including one reached through an approval; one call per
  thread, claimed atomically when the socket opens; a vd_ id never bypasses the Live
  call guard; 30 utterances/min per person.
- Speech providers behind services/speech (Gemini REST STT/TTS, ElevenLabs), chosen in
  Settings (GET/PUT /api/settings/direct-voice, admin, audit-logged).
- The #2694 delta cursor treats a Direct reply as the agent's own session answering.

Measured end to end with real speech: first spoken word 3.4-4.4 s after send on Claude
Code, 10.7-12.0 s on Gemini CLI; STT 1.1-1.9 s.

The speech providers and the bridge were written before their tests. Every gate was
then mutation-checked: 16 backend gates go red when removed, run with bytecode caching
off.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* feat(workspace): two call buttons, and the Live orb in the Direct call's strip (trinity-enterprise#1046)

Operator review of the first build:

- The Call control's mode menu is gone. The composer has two buttons, one per
  call: the Live call (waveform) and the Direct call (handset). Each carries its
  own reason when it is off; while one call is on, its button ends it and the
  other says "Another call is on". Nothing is remembered per browser any more
  (`CALL_MODE_KEY`, `loadCallMode`, `saveCallMode` and `CallModeMenu.vue` are
  removed; `callButtons` is the pure rule).
- The Direct strip's flat pulsing disc is replaced by the Live call's orb in a
  48px disc. The overlay's smoke and core move into `chat/voiceOrbField.js`
  (one renderer, an instance per orb, no module state) and `VoiceOverlay.vue`
  draws through it unchanged. `chat/VoiceOrb.vue` draws the same field clipped
  to a circle on the orb's black ground, so it reads in both themes: 80
  particles, a DPR-capped bitmap, the first frame drawn on mount, one still
  frame under reduced motion, and the core tinted with the state — green
  listening, amber working, blue speaking — because at this size the smoke is
  a rim and cannot carry it. `directOrbState` maps the Direct statuses onto the
  orb's vocabulary.
- While the person talks, the orb moves with their voice: the composable feeds
  the mic RMS into `amplitude` whenever nothing is playing (`rmsOf` is exported
  from the endpointer).

PortalConversation's raw_gray count stays at its baseline: the two buttons
share one class recipe (#2662 arms) instead of repeating it.

Tests: `portalDirectCall.spec.js` covers the button rule, the status→orb map and
the mounted orb (draws, animates, stops on unmount, still under reduced motion);
`portalDirectCall.mount.spec.js` drives both buttons in the real conversation.
Verified in the browser on the local stack, light and dark, Live and Direct.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

* feat(workspace): the Direct call's strip shows four call bars, not a canvas orb (trinity-enterprise#1046)

The strip's mark is now `portal/CallBars.vue`: four rounded bars in a disc
tinted with the status tokens, sitting on the theme's own surface. Green bars
follow the person's mic while it listens, amber bars run a slow wave while the
turn is worked on, blue bars follow the agent's voice while it speaks, and
muted is four flat gray dots. Four elements and no canvas; under
prefers-reduced-motion it holds one still frame per state.

The canvas orb and the field extracted to share with it are gone, so the Live
call's VoiceOverlay.vue is back to dev's version and this PR no longer touches
it.

Tests mount the bars. Each of three mutations turns a test red: ignoring the
sound, ignoring mute, and starting the loop under reduced motion.
`portalUndefinedCalls` now lists `cancelAnimationFrame` as a browser global,
next to `requestAnimationFrame`.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(workspace): a Direct call answers a reply it cannot read aloud, and its strip is one quiet row (trinity-enterprise#1046)

A reply that is only code, a table or links had no speakable sentence, so the
call said nothing and looked dead. test-echo's fenced echo hit this on every
turn. A successful reply with nothing to read aloud now gets "The reply is in
the chat." (LINE_REPLY_IN_CHAT); a reply with any spoken sentence does not.

The strip is now one fixed-height row, so the thread below it no longer moves
when the state changes: 32px bars, one short word (Listening, Got it,
Thinking, Speaking, Yes or no?), one detail line (a notice, what was heard,
the tool at work, or the slow-runtime wait), a mic icon that shows a slash
when muted, and a red hang-up handset. The "Direct call · runtime · speech
by …" line and the first-word timing are gone; Esc moves to the hang-up's
tooltip.

Tests written red first: two backend (the in-chat line, and no line when
something was spoken) and four frontend (the detail order, short words, and
the mounted row with icon buttons and a danger-token hang-up).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(workspace): a Direct call reads a reply that is only a code block, instead of pointing at the chat (trinity-enterprise#1046)

test-echo answers with nothing but a fenced block. The previous fix spoke
"The reply is in the chat." for it, which is no answer. When a reply has no
prose outside its code blocks, the text inside the blocks is the reply and is
now read (fence markers and the language tag dropped). Code beside prose
stays unsaid. The in-chat line remains only for a reply with nothing to read
at all, such as a table alone. The rule lives in spoken_text.speakable, so the
streamed speech and the reconciled tail agree and nothing is said twice.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Revert "fix(workspace): a Direct call reads a reply that is only a code block, instead of pointing at the chat (trinity-enterprise#1046)"

This reverts commit c214c9e4d. Operator ruling: a reply that is only a code
block is correctly answered with "The reply is in the chat." Code is not read
aloud, whether or not prose sits beside it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(voice): a Direct call speaks an execution once (trinity-enterprise#1046)

Each speak frame built a fresh SpokenBudget, and only utterances were
rate-limited, so re-sending speak for the same run bought its synthesis
again, up to the call's 30 minutes and across calls. The bridge now claims
an execution before speaking it: Redis SET NX across calls and workers,
plus a per-socket set that holds when Redis cannot answer.

Also pins the store<->backend contract of a Direct call (mode on start,
voice_call_id on a spoken turn, the roster's direct_voice), each of which
could be deleted with every test green.

merge-train 2026-11-02, mechanical per operator ruling at Gate 1.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(voice): stub the Live capability in the one-call-per-thread test (trinity-enterprise#1046)

The live case reached realtime_voice_capability before the cross-kind 409
and got a 503 on a host with no realtime provider key (CI). merge-train.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: trinity-ability <309458136+trinity-ability@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* chore(enterprise): bump submodule — invite door and terms acceptance (trinity-enterprise#1054) (#4200)

src/backend/enterprise 47cb75a → 452ea04 (abilityai/trinity-enterprise#1074).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* feat(auth): a grant's use leaves an audit row, and grant changes share one verb (abilityai/trinity-enterprise#871) (#4174)

An agent key a grant admits now writes one `capability_used` audit row: the
agent as actor, its key id, the capability, the target, the method and the
endpoint. It is written inside `enforce_agent_capability`, after the owner
bound, so it means the grant cleared the call, not that the change landed, and
a refused call writes only `capability_refused` (a holder aiming outside its
owner's agents included, code `capability_target_not_owned`, same 404). People,
`user` and `system` keys write none. One row per capability and target per
request; manifest deploys and an honoured skill-set probe record their own use.

`skills.manage` grant changes log `capability_grant` / `capability_revoke`
through every door; earlier rows keep `skill_manager_*`.

Security fixes found by `/cso --diff` and review
(docs/security-reports/cso-diff-2026-11-02-871-capability-used.md):
- `owner_owns_agent` now refuses the system agent. Owned by the admin, it was
  reachable by any holder an admin owns, and what lands there runs under a key
  that bypasses every permission check.
- The file routes apply the owner bound, so an edge to another user's agent no
  longer lets a holder write its instructions or skills (now 404).
- The skill-gate map read uses the same bound.
- The create row names the slug the agent would get, not the raw body.

Also: `request_id` on the agent `delete` row; `test_ent1030_gates.py` no longer
reaches the host's Docker. Out of scope: issue item 3 (outcome rows for schedule
writes and the timeout PUT); `trinity_prompt` via create-mode deploy stays #4025.

Fixes abilityai/trinity-enterprise#871

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(autonomy): declare get_seat_autonomy in the self-gate census; count get_autonomy in the MCP docs (ent#641)

Fallout from merging dev:
- test_ent719_self_gate_identity drives every assert_own_agent route with a
  well-formed request. The new GET /seat-autonomy needs an execution_id query
  parameter, like its seat-decision siblings.
- tool-count-docs (ent#991 T5) reads the registered total from the user docs.
  dev's search_memory plus this PR's get_autonomy make 177 tools (172
  operator), and decisions.ts now has 3 tools.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(gemini): dev's wall-clock and usage-statistics tests run as the owner's turn (ent#926)

test_ent1003_gemini_wall_clock and test_ent986_gemini_usage_statistics came in
with the dev merge and drive the Gemini runtime with an execution id. This
branch's tier check asks the platform whose turn it is, and with no platform
to ask it fails closed (503). These tests are about the CLI, not the turn, so
they take the same owner-turn stub the branch's other Gemini tests use
(turn_scope.fetch_turn).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* feat(skills): find a skill and see what Save would do, on bounded lists (abilityai/trinity-enterprise#660) (#4178)

- Assign skills dialog: text, source and category filters narrow the rows
  and both bulk buttons, never the draft. Each row and a footer count say
  what Save would do: "to add", "to remove", or "kept by set" for a skill an
  assigned set still holds. The list keeps one height while the filter
  narrows, in a panel capped to the window.
- Library and the agent's Skills tab: the lists scroll in a BoundedRegion
  that fills the window, so the page no longer grows with the library. The
  Library gets the same filters and a stated total. On the tab, Own and
  Shared share one box, a total sits in the header, and Shared's heading
  docks to the box from 640 px up.
- BoundedRegion is now `relative`: unpositioned, the cards' sr-only labels
  escaped the box and Chrome counted them in the page height.
- Backend: `category` on SkillInfo, from the skill's frontmatter. No
  migration, no new route, no base-image or MCP change.
- e2e/skills-bounded-lists.spec.js (Chromium, @smoke, hermetic) pins the
  geometry that jsdom and Firefox cannot see.

ent#660 was re-scoped because its text predates ent#754. Edition:
OSS-core, ungated (confirmed on ent#660).

Fixes abilityai/trinity-enterprise#660
Fixes #3448

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(access): declare get_seat_autonomy's tier in the agent route census (#4206)

#2984 added GET /api/agents/{agent_name}/seat-autonomy and #3778 added the
census that requires every agent-naming route to declare a tier. They merged
one after the other without seeing each other, so dev fails
test_ent926_agent_route_census::test_every_agent_route_is_declared.

Declared accessor, with a reason, matching its sibling decision routes'
gates (assert_own_agent, then assert_agent_access).

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* fix(slack): fall back to Public URL for OAuth callback redirect (#4167)

Motivation: with FRONTEND_URL empty (the .env.example default), the Slack
OAuth callback redirect fell straight to http://localhost, so after
Install to Workspace the admin landed on a dead page even though the install
succeeded. The redirect URI for the same flow is already built from the
Public URL setting.

Approach: get_oauth_callback_redirect now uses FRONTEND_URL if set (unchanged
precedence), else the Public URL setting, and only then the localhost
fallback with the existing error log. This covers the platform and per-agent
success and error redirects, since they share the function.

Validation: ran tests/unit/test_slack_*.py plus the new
tests/unit/test_3563_slack_oauth_redirect_base.py: 227 passed. With the fix
reverted, te…
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants