Skip to content

v1.0.5 — Projects you run work from, agents that wake, searchable memory

Latest

Choose a tag to compare

@vybe vybe released this 17 Nov 13:47
· 1 commit to main since this release
87e21ff

v1.0.5

Released: 2026-11-17
Tag: v1.0.5  ·  Diff: v1.0.0...v1.0.5

478 commits · 361 merged PRs · 2,129 files (+300k/−21k) · 162 public + 128 private-tracker issues (109 listed below) · enterprise submodule 982735f → a26e522 · release candidates rc1–rc4 cut from dev along the way; the release is cut from the dev tip, two commits after rc4 (a user-docs wording fix and the main reconcile merge).

No freeze plan was written for this cycle. User-facing summary: What's New in v1.0.5.

Highlights

1.0 put an agent in a seat on the team. 1.0.5 makes the seat workable day to day: projects
people and agents run work from, agents that wake when the thing they were waiting for
happens, a memory that can be searched, and an install that is closed to the network until
an admin opens it.

  • Projects are a working environment. A project has a board view, tasks with subtasks,
    waits-on and an editable title, objective and definition of done, a Runs tab listing the
    executions behind it with cost and queue age, and context an agent reads (a written or
    uploaded brief). A task's run starts from the project. A steward agent is picked at
    creation, proposes members and task owners for one approval, and works in a per-project
    mode — Track, Propose or Auto — that the platform enforces. An existing GitHub epic can be
    imported as a project, and a task can point at its issue or pull request (ent#957,
    ent#1232, ent#1133, ent#1105, ent#1193, ent#1195, ent#1169, ent#980, ent#1235, ent#1236,
    ent#1163, ent#1240, ent#867).
  • Agents wake on their own. An agent is woken when an ask it raised is answered, and a
    caller is woken when the run it delegated ends or asks — both on by default, with a missed
    wake delivered on the next start and a per-hour budget on caller wakes. Platform-started
    turns run on the agent's configured model (ent#790, ent#1108, ent#1102, #4357, #4536,
    #4442).
  • Asks reach someone who can answer. An ask goes to a person who can open the agent, an
    ask raised in a room comes back to that room, an agent on a project can address the
    project's people, a changed primary re-addresses what is pending, and a refusal names who
    a card is waiting for (ent#1135, ent#1134, ent#1095, ent#1096, ent#1042, ent#1053, #4501,
    #4579).
  • Retrieval memory (default off). Turns are captured, recalled into the prompt and
    searchable by the owner and by the agent over MCP (search_memory). Agent-written memory
    about a person passes anti-poisoning gates, and the owner sees a person's memory history
    with undo (ent#1022, ent#1027, ent#1028, ent#419, ent#962).
  • Voice. A Direct call sends every spoken turn to the agent and plays its own reply,
    with stop by voice or by hand; send_voice_reply can address a person by role or email so
    a scheduled run can speak (ent#1046, ent#1087 in part, ent#700).
  • Systems reconcile. apply_system reconciles a live instance to its manifest instead
    of suffixing _2, and a manifest states the whole target: skill assignments, admin
    grants, the approval map, role ids and catalog-only members (ent#695, ent#1030, ent#991,
    #4128–#4132, #4151).
  • Closed by default on the network. External MCP access is off on a fresh install and
    switched on in Settings, agent SSH ports publish on loopback, hosted installs trust only
    their own proxies for the client address, and DO_NOT_TRACK is one switch after which the
    instance contacts no vendor host. Agent runtimes stop sending usage fingerprints to their
    vendors by default (ent#1101, ent#1168, ent#873, #3047, #4499, ent#986, ent#1093).
  • Access tiers tightened. Execution reads and agent control verbs moved off the
    accessor tier, the accessor tier is no longer the default for agent routes, a suspended
    account is refused at every door and loses its live connections, and peer skill grants are
    enforced inside the container (ent#875, ent#926, ent#1258, ent#863, ent#925).
  • Operating an install. trinity update is one command with preflight, published-URL
    verification and rollback; the admin password can be changed in the app; every compose
    file defaults to a bundled PostgreSQL, with SQLite still available behind an explicit
    setting; an alert names the running agents still on an old base
    image (#3007, #3825, #4175 in part, #4479, #4483).
  • Funding an agent with a card. An operator enrolls a card and creates a mandate for an
    agent from the Payments tab, backed by Nevermined widget sessions and per-agent delegation
    binding (ent#966, ent#967; ent#965 stays open for step 3).
  • Instance-to-instance. Cross-instance A2A signs in with AAuth agent identity, and a
    credential ask can name an agent on another Trinity as its destination (ent#862, #4262,
    #4401).

Features

  • #3007 feat(update): a self-hosted operator cannot complete an upgrade from the documented path — make trinity update a first-class command with preflight, published-URL verification and rollback
  • #3871 Opt-in PDF uploads — let operators deliver PDF originals to agents through the existing upload path (off by default)

Private tracker (trinity-enterprise):

  • ent#225 Rename agent "Sharing" tab to "Channels" (IA reconsideration)

  • ent#450 Access tab: suggest registered Trinity users when adding an operator by email

  • ent#459 feature: Access tab — send & re-send a pending operator's access invitation

  • ent#471 Raise the default agent turn limit (max_turns) from 50 to 200

  • ent#529 feat(skills): the seeded default community source follows the platform's bundled tag pin on upgrade

  • ent#539 Skills tab: Select all / Deselect all for batch skill assignment

  • ent#584 Capability descriptions: one source, every surface — Workspace Info tab, agent Info tab, A2A card and typeahead disagree because they read five different fields

  • ent#660 Skills tab: redesign the assignment layout — find, select and de-select skills in one place

  • ent#695 feat(systems): deploy_system apply mode — reconcile existing members and schedules instead of suffixing _2

  • ent#700 feat(voice): proactive voice notes — give send_voice_reply an addressing surface so a scheduled run can speak (today it is channel-turn-only)

  • ent#701 Workspace rooms: seat a second internal person as a participant — the seat half of #630 (Enterprise)

  • ent#736 Deprecated library skills are not marked on skill-set members, set-assign notes, the skill-runner picker or the CLI table

  • ent#790 Wake an agent by default when an ask it raised is answered (flip operator_resume_enabled to on)

  • ent#867 Projects: a task's owner can't be changed after the task is created

  • ent#934 Telegram: an album (several photos sent together) arrives as N separate agent turns — coalesce on media_group_id

  • ent#951 Settings › Credentials: its own pane in the agent Settings rail (not a section of Storage)

  • ent#957 Workspace Projects: Kanban board view of a project's tasks — one column per status, drag to move

  • ent#959 Workspace Inbox: room messages show under Unread and All (split from ent#610 AC 2)

  • ent#964 Create Agent: prefill the name and ID from the repository or template name (without the owner)

  • ent#966 feat(payments): Nevermined widget sessions + per-agent card delegation binding — backend, API, MCP (ent#965 step 1)

  • ent#967 feat(ui): Nevermined spending panel — enroll a card and create a mandate for an agent inside Trinity (ent#965 step 2)

  • ent#968 Workspace agent band: replace the stacked activity bars with a waveform signature (Rhythm)

  • ent#981 Projects: a task reaches its owner from the project record, not through an agent

  • ent#991 Read permission edges and self-change grants over MCP, so no target-view row stays unobserved

  • ent#1006 Declare and edit an agent's metrics from the UI — form writes the template's metrics: block, template stays the one source

  • ent#1022 feat(memory): retrieval memory layer — mem0 OSS embedded in the backend, captured at the turn seams, recalled into the prompt (default off)

  • ent#1024 Admin user MCP key can act on behalf of an agent — agent_name on ask_operator, project, report and metrics tools

  • ent#1027 Retrieval memory: extract headless inside the agent container on the agent's own runtime/subscription

  • ent#1028 Retrieval memory: search — owner route, agent door + MCP search_memory tool, search box in the Memory panel

  • ent#1030 SystemManifest carries the whole target state: skill assignments, admin grants, approval map, role ids, catalog-only members; apply_system honours them

  • ent#1042 An answer to a role-addressed ask is not limited to the people seated in that role

  • ent#1046 Workspace voice: a Direct call mode — every spoken turn goes to the agent and you hear its own reply

  • ent#1102 Wake on ending by default: a playbook call wakes its caller when the callee ends or asks, and an autonomy-off person-facing agent has a wake path

  • ent#1105 Project-scoped execution view: a Workspace Project lists the executions behind it across its member agents, with cost and queue age

  • ent#1107 Terms of use per instance: a shipped default, admin-editable, shown once to every first-time person on any door

  • ent#1108 Agents wake on an answered ask by default, on every instance

  • ent#1131 Projects: a room linked to a project doesn't say its agents aren't on it, and a member can't ask to add them

  • ent#1132 Projects: a viewer gets no explanation for missing controls and can't ask to join

  • ent#1133 Projects: a person can't edit a task's title, objective or definition of done

  • ent#1134 Asks: an agent on a project can't reach the project's people, and a fallback to the operators is invisible in the room

  • ent#1135 Rooms: the answer to an ask raised in a room never comes back to that room

  • ent#1153 People picker primitive: BasePeopleInput form field + one scoped people-search seam, extracted from #4306

  • ent#1154 Agent assignments: pick the person and the new primary with the people picker, usable by a non-admin owner

  • ent#1155 Projects: members, guests, steward and task owner pick a person instead of typing an email

  • ent#1161 Projects: a person can't take part in a project until they hold a seat on the stewarding agent, and the only lever offered is making them a collaborat

  • ent#1163 Projects: turn an existing tracker epic into a Workspace Project

  • ent#1169 Projects: pick a steward agent at creation; the steward carries the library's project-management skill set

  • ent#1193 Projects: prepare a task's context on the project — write or upload a brief, mark what agents get (ent#1136 step A)

  • ent#1195 Projects: start a task's run from the project — pick the agent, see what it gets, open the task's chat (ent#1136 step B)

  • ent#1232 Projects: a task can't have subtasks, can't wait on another task, and can't be "not started"

  • ent#1236 Projects: a steward agent runs only if someone remembers to schedule it

  • ent#1237 Agents have no standard way to keep and read files on the platform

  • ent#1240 Projects: a task can't point at the issue or pull request where its work happens

  • ent#122 Epic: Fresh-install fleet provisioning — manifest-driven system install without a PAT, UI manifest install, fork-to-own, guided agent credentials

Fixes

  • #1988 CANARY_ENABLED silently disables on any truthy value except "1"
  • #2328 bug(dispatch): per-task max_turns is write-only on the push path — stored, restored, never sent to the agent
  • #2748 fix(mcp): create_room advertises default budgets 60/24; backend applies 200/168
  • #2749 fix(ui): chat file picker omits .zip although uploads accept ZIP
  • #2891 bug: an installed pytest-randomly silently randomizes local runs, surfacing latent order dependencies as phantom regressions
  • #3009 bug: DigitalOcean 1-Click can be deployed on plans below the 4 GB / 2 vCPU minimum, and runs out of memory on 1 GB
  • #3060 bug: Workspace landing layout shifts: rail column jump, chat-shaped skeleton, late tab badges
  • #3089 fix(metrics): a negative stored metrics_daily_point_cap reads as unlimited
  • #3136 Assigning a skill reports "delivered — available now" while dropping the injection's missing-dependency warnings
  • #3254 bug: the remaining platform alert emitters still file one row per reading — move them onto the shared alert path (#3246 part 2)
  • #3296 bug: an identical delegation from a later turn replays the first run, so report-back goes to the first turn's conversation
  • #3298 bug: an /chat or /task idempotency claim left in_flight by a dead process answers 409 for 24 h
  • #3300 fix(mcp): a broad list_reports under an agent key filters after the backend's limit
  • #3302 fix(ui): the Operations queue windows truncate silently past their row limit
  • #3317 bug: reconcile_after_restart racing a live loop advance runs the same iteration twice
  • #3363 bug: scheduler retry lacks cron-fire parity — skips Workspace delivery and the autonomy-off / sync-freeze gates; three trigger lists omit retry
  • #3380 bug: a watchdog-recovered chat turn never writes its assistant message — row says SUCCESS, thread shows nothing
  • #3390 bug(tests): 40 unit tests fail only in sequential full runs — bisect the order-dependent polluter
  • #3395 bug(tests): schema parity is one-directional — a forgotten MIGRATIONS line ships green and breaks writes on existing SQLite installs
  • #3408 Agent runtime: process-ownership fix (#3308) covers Claude chat and headless only — Codex/Gemini keep a stale registry entry, and a failure before the reader starts leaks one
  • #3439 bug(workspace): a completion report written into the open thread is not shown until reload
  • #3444 bug(ui): Login has no return path: expired session and deep links always land on the dashboard
  • #3445 bug(ui): Header connection indicator stays Disconnected after a client-side login until the page is reloaded
  • #3448 bug(ui): Skill lists render the whole library unbounded with no search or filter
  • #3451 bug(ui): Operations lists grow the page instead of scrolling in a bounded viewport
  • #3466 bug(ui): An operator whose session ends in the Workspace gets the client email-code sign-in with no route back to the operator login
  • #3487 bug: a Workspace turn on a pull-pilot agent does not walk the subscription pool (SUB-003 walk is push-only)
  • #3495 Rooms: platform members can set room budgets past the operator caps
  • #3542 Agent page at phone width: the avatar circle overlaps the left edge of the header card
  • #3543 bug: messages point to "Settings → GitHub token", which does not exist — the personal token card is on the MCP Keys tab
  • #3562 Projects list: two primary "New project" buttons when there are no projects
  • #3563 bug(slack): after Install to Workspace the browser is redirected to http://localhost when FRONTEND_URL is empty
  • #3565 New project dialog changes height when "Who can see it" is switched
  • #3566 Project Overview: the Agents section is blank when the project has no agents
  • #3568 Relative times read "1 minutes ago", "1 hours ago", "1 days ago"
  • #3569 Project Members & access: the agent picker goes blank after adding an agent
  • #3584 Public links: saving the Edit dialog shifts the expiry by the browser's UTC offset, and an expiry can't be removed
  • #3601 Webhook URL is shown as http:// behind an HTTPS proxy
  • #3603 A failed channel turn posts the raw platform error into the chat (in a Telegram group, every member sees it)
  • #3618 bug(pull): Stop on a pull-pilot run records "Execution error: Claude Code reported no error detail (diagnostic: result_type=user …)" instead of "Execution terminated by user"
  • #3625 bug(pull): every task on a pull pilot is reported as "Agent at capacity — queued", even when the agent is idle
  • #3626 bug(tasks): the Tasks form Timeout dropdown offers up to "1 hour" and silently clamps to the agent's execution timeout — the cap is never shown
  • #3651 bug(activities): direct-dispatch companion activities leak on recovery writers beyond the watchdog claim; a2a queued-cancel sites call no closer (#1804 class)
  • #3663 Access tab at 375 px: the "Remove" link on a person's row is cut off
  • #3664 Access tab: a stale "Shared with …" note and a stale platform-role badge
  • #3691 Flaky: test_ent751_gate_callers TestPortal held-turn notice is not persisted ~1 in 8 runs (order-dependent)
  • #3706 bug(rooms): a person's message over 8,000 characters is refused, while 1:1 chat accepts 100,000
  • #3709 Workspace: a gated request's outcome posted into an open thread shows only after a reload
  • #3727 Own-repo agents commit .claude/policy-limits.json.stamp.json on their first auto-sync (missing from the managed .gitignore)
  • #3821 Private-network plain HTTP is unreachable by IP: the on-demand wildcard site captures dotted-IP hosts
  • #3838 bug(codex): a prompt over 128 KiB fails to spawn — passed as one command-line argument
  • #3849 bug: a prompt-length rejection is classified as an auth failure and walks every subscription (SUB-003)
  • #3868 Agents cannot write their own .env.example — the .env.* protected path also blocks the names-only template the compatibility check requires
  • #3872 fix(ui): the agent header jumps while the GitHub line loads
  • #3879 dev-nightly red: Tier 3 failed on dev
  • #3881 bug: a Public URL with a trailing dot is accepted, and Telegram webhooks then fail TLS silently
  • #3884 bug(setup): password meter shows "Strong" while the 12-character minimum is unmet, and the create button is silently disabled
  • #3894 [BUG] Owner can't reach their own UI-created agent in the Workspace: not in the roster from the platform click-through, and People rejects self-share
  • #3922 bug(workspace): the same platform account gets the client view when it signs in on the Workspace's code page — no Projects, none of its own agents
  • #3932 bug: pull-claimed executions never open a chat_start activity — timeline is blank on pull agents
  • #3940 Scheduler spins and leaks memory without bound at DST fall-back: CronTrigger next fire time goes backwards for */N crons in non-UTC zones
  • #3944 fix(mcp): fan_out idempotency key ignores the caller's turn (#3296 follow-up)
  • #3949 A peer dispatch that only MENTIONS a library skill with a leading slash is refused 403 skill_not_granted — the match-anywhere rule is terminal on the ent#662 path
  • #3950 fleet-health reports 0 unhealthy / 0 unknown while most agents have no health row — the buckets omit row-less agents, and stale verdicts are served as current
  • #3966 bug(scheduler): the pre-check hook can't tell which schedule is firing, so an agent with two cron schedules runs the wrong task
  • #3982 fix(pull): terminate landing between claim and activity open leaves a started bar open
  • #3984 fix(sub-003): a turn-limit 422 from the agent is read as a credential failure when the turn cap is 401 or 403
  • #3988 fix(task): a turn-limit failure answers sync /task callers 503 "agent may be unavailable" with no error code
  • #3989 bug(canary): S-01 fires on every running /chat turn — the slot is keyed on a uuid that is not the execution row id
  • #3993 bug(scheduler): max_turns failures are retried though every retry repeats them
  • #3994 fix(workspace): a project's Needs you row overflows its card when the agent name, kind and time don't fit
  • #3998 bug(chat): a /chat turn whose setup raises leaks its capacity slot and idempotency claim
  • #4000 bug: git maintenance and git-health stats only run inside auto-sync — self-pushing agents bloat with no repack and no alert
  • #4001 bug: git status poll's 30 s fetch starts git auto-maintenance and kills it, stranding maintenance.lock
  • #4002 bug: orphan sweeper kills ~/.trinity/setup.sh after 90 s, and TMPDIR on the workspace volume is never cleaned
  • #4027 bug(mcp): deploy_system throws on a total-failure deploy instead of returning the report
  • #4048 bug(alerts): a person's ending of a skills alarm silences a worse reading for 7 days
  • #4050 bug(alerts): the platform-alert seam can report created for a reading it did not file (two kinds sharing an id prefix and key in one microsecond)
  • #4051 test: every xdist worker in the unit tier shares one SQLite file, contrary to the CI comment that says each worker has its own
  • #4052 test: test_2610_room_jsonl_keepset errors when it runs after a file that puts tests/unit on sys.path
  • #4053 bug(alerts): the platform can still end a must-act alert at runtime; only a static scan forbids it
  • #4055 bug(alerts): a failure inside a task the platform-alert seam spawns is never logged
  • #4057 bug(loops): a loop dispatched from a park skips the post-run decision of the run before it
  • #4058 bug(loops): run rows duplicated before the #3317 fix can strand reconcile
  • #4060 bug: MCP clients stay disconnected after every redeploy — endpoint is down longer than the client's reconnect window
  • #4064 bug(gemini): the Gemini child is spawned without a process group, so terminate() resolves to the agent server's own group
  • #4072 bug(workspace): the inbox-files notice reads as a fresh upload on every turn, so agents act on files from other conversations
  • #4085 bug: scheduler-created executions (cron, reminders) never record subscription_id — SUB-004 spend misses them
  • #4093 A BaseToggle label that wraps sits centred beside its switch
  • #4105 bug(pull): self_task and skill_gate are in neither pull trigger set — pushed on pilots, overbooking invisible to M3
  • #4123 bug(sub-003): a real auth failure mentioning "prompt is too long" does not walk subscriptions, and the turn-limit anchor is unpinned
  • #4128 System detail and export lose permissions and folders
  • #4129 GET /api/systems omits systems whose members have bare names
  • #4130 Apply and deploy results misstate what happened in five edge cases
  • #4131 naming: bare edge cases: re-parse, name collisions and restart scope
  • #4132 A system export cannot always be applied back unchanged
  • #4137 test(memory): the memory_on fixtures don't stub last_error, so a stale embedder error fails them
  • #4149 start.sh --hosted rewrites the Caddyfile and drops operator-added site blocks on upgrade
  • #4151 apply_system drops a schedule's model field without a warning
  • #4157 bug(workspace): a platform account that starts at the Workspace sign-in needs two emailed codes
  • #4160 bug: Inbox Unread — scrolling a room's reading pane scrolls the whole Workspace page
  • #4165 bug: github:owner/repo@ref drops a dotted tag and clones the default branch
  • #4228 fix(voice): send_voice_reply does not scrub staged credential values before synthesis
  • #4229 bug: Workspace Inbox ask row jumps in height when answered — badge line is empty while pending
  • #4241 Agent page: the last tab paints over the pinned Settings tab ("FilesSettings")
  • #4256 test: test_2742 lock-free status control intermittently loses its teeth in CI (sampler misses index.lock)
  • #4257 bug(dispatch): async /task dispatches to an idle pull-pilot agent sit queued for hours and start in batches when another slot is released
  • #4259 bug: cleanup sweep revokes a just-minted agent MCP key while the container is still starting
  • #4280 classify-submodule-failure.sh misclassifies failures under pipefail (SIGPIPE race with grep -q)
  • #4302 bug(rooms): a stopped mention chain is a dead end, with no plain explanation and no way to continue
  • #4331 systems: manifest name and member-key checks accept a trailing newline
  • #4349 Workspace: a chat with a new unread reply stays hidden under the tab strip's "N more"
  • #4352 fix(tasks): the execution error's reader phrases are matched case-sensitively, so the portal loses its 504/429 on mixed-case text
  • #4353 Workspace: the sidebar New chat button duplicates the roster click and the header New chat
  • #4357 bug(delegation): wake-on-ending and reminder turns inherit the chain depth, so an autonomous delegate→wake→delegate run stops after ~8 rounds
  • #4369 bug: webhook context is framed but never forwarded — the agent never sees it
  • #4381 trinity_version always reports "unknown": operator-intake and export read GIT_COMMIT_SHORT, which nothing sets
  • #4442 bug: platform-started turns (wakes, reminders, operator answers) run on a different model than the agent's configured one
  • #4445 test_ent1117 wildcard port-probe test flakes with EADDRINUSE on CI
  • #4470 bug(git): a manual Pull moves a tag-pinned agent onto the default branch
  • #4479 SQLite→PostgreSQL guide and start.sh refusal: image step is incomplete (build backend only → mixed-version stack; rollback names no rebuild)
  • #4483 bug(upgrade): agents stay on the old base image after an upgrade and nothing says which ones — agent-side fixes are silently inert
  • #4501 bug(asks): an ask to: approver is filed for a seated approver who cannot open the agent, and then nobody can answer it
  • #4505 Room: a short gated-skill request is refused as too long when another participant's long message is unread (v1.0.5-rc2, B1)
  • #4515 Workspace: after accepting the terms, the roster never reloads and says "Accept the current terms to continue"
  • #4536 bug(wake): a fast callee's ending is dropped by the 15 s wake window, so a delegate→wake loop stalls silently
  • #4541 bug(deploy): start.sh --hosted upgrade checks the PROXY listener before the new web container answers, so Caddy stays on the plain listener and every visitor shares one address
  • #4570 GitHub template bundled avatar is installed but never shown (avatar_url stays null)
  • #4579 bug(asks): a credential ask is filed for an assigned primary who cannot open the agent, and then nobody can fill it
  • #4597 bug: Project page header squeezes the project goal into a narrow column beside the action buttons

Private tracker (trinity-enterprise):

  • ent#1010 Workspace Inbox: a message that lands between the last fetch and the read write is marked read unseen (chats and rooms)
  • ent#1011 Rooms: one email seated under both identity spaces shares one read cursor
  • ent#1082 First sign-in of an invited member can land in a dead Workspace instead of the terms screen
  • ent#1091 Task-reach alerts flood the owner's Inbox: one row per task per event, stale filed high, no daily roll-up (ent#981)
  • ent#1095 Changing or first setting an agent's primary person must re-address its pending asks and gated approvals to the new primary
  • ent#1096 A pending approval cannot be answered after the agent's primary changes: still bound to the old addressee, no admin override, and the refusal does not name who
  • ent#1216 bug(executions): agent-key execution reads return 404 / empty since 2026-11-08 ~21:58Z — receipts unreadable while running, results cut, and an unfindable sync timeout produced a duplicate run
  • ent#1221 An agent cannot message its primary when the primary is a Workspace-only member: send_message returns recipient_not_found
  • ent#1255 Invite emails link to the public chat host when FRONTEND_URL is unset (/login 403)

Security

  • #2530 docs: the setup.sh convention never says "no secrets" — agents are persisting GITHUB_TOKEN into ~/.profile
  • #3047 bug: per-IP rate limits likely see one client IP on Caddy-fronted (hosted) installs
  • #3180 feat: a bounded, attributable machine credential for operational WRITES that survives enforced 2FA (write tier for the ops key scope)
  • #3322 bug: URL validation lets raw UnicodeError / ValueError escape, so some bad URLs give a 500 instead of a named 400
  • #3402 bug(gates): a held gated message is replayed to the agent in later history — it can act on a never-approved request
  • #3549 bug(audit): a permissions replacement is audited with no targets — permissions_set details.targets is always null
  • #3799 DigitalOcean installer puts the admin password in droplet user-data
  • #3825 No way to change the admin password after first-run setup — not in the API, not in the UI
  • #3886 bug(security): agent base image ships EOL Node.js 20 — unpatched against the Oct 2026 Node CVEs (Brotli OOB write, TLS hostname bypass)
  • #3909 bug: an MCP URL with an unencodable host escapes mcp_validator as a raw UnicodeError (likely 500)
  • #4025 bug(systems): create-mode deploy lets a creator or an agent key overwrite the platform-wide trinity_prompt
  • #4026 bug(mcp): a returned operational failure is audited as a successful tool call
  • #4121 write_user_memory accepts a finished run's id, so an agent key names another person's seat run and writes that person's notes
  • #4156 bug: an agent can say "Saved" while the note goes to shared auto-memory, not write_user_memory
  • #4284 bug: exception text in HTTPException detail across backend and agent-server routes (py/stack-trace-exposure siblings CodeQL does not flag)
  • #4359 fix(voice): Workspace read-aloud sends unscrubbed text to TTS, and the direct-call scrub is not pinned by the parity guard
  • #4435 bug: dev frontend image deletes package-lock.json and runs npm install — every dev build floats the whole dependency tree
  • #4475 Terminating a running execution writes no audit row when the in-flight request wins the status race (v1.0.5-rc2, B20)
  • #4482 Platform-started turns report raw provider failure text and subscription names into a person's Workspace thread
  • #4499 Privacy-strict installs: two outbound channels have no off switch, and the bug-report switch does not work on the published images

Private tracker (trinity-enterprise):

  • ent#21 feature: Per-client access controls (block / revoke / approve) for channel users
  • ent#341 feat: per-agent tool-access profiles — 3-4 named presets with Read-Only as one of them (human-only, no agent MCP surface)
  • ent#419 feat(memory): anti-poisoning gates on agent-written user memory (GS-3 + PV-7)
  • ent#591 Agent creation is an ambient grant — make it a declared capability, optionally bounded to a catalog
  • ent#595 security: the voice WebSocket authenticates with the platform JWT in the query string — the bearer token lands in proxy access logs on every call
  • ent#643 Default prompt injection: instruct agents not to shell out to the claude CLI — credentials are turn-scoped by design
  • ent#717 Admin-granting whitelist adds and role changes accept key-authenticated admin principals — a minter weaker than the session it produces
  • ent#719 Owner-tier grant routes an agent or system key still reaches — classify or gate the human-only route baseline (follow-up to #2996 / #711)
  • ent#737 Channel access requests capture the requester's channel identity — the owner approves a person, not an email
  • ent#780 security: the agent's own readback (get_my_ask) returns operator-queue rows about a person — the fourth read the #715 fix missed
  • ent#781 security: an admin-written evaluation names its grader (admin:) to the graded agent and shared readers
  • ent#822 File routes: a symlink in an agent's home passes the lexical path checks, and the agent server follows it
  • ent#830 bug: idempotency scope for agent trigger boundaries is not bound to the caller, so callers can replay each other's responses
  • ent#840 security: make the agent-key self-gate a route-census policy class — eight modules carry it by hand, nothing enforces it
  • ent#862 Cross-instance A2A signs in with AAuth agent identity (graduate the prototype for Trinity-to-Trinity calls)
  • ent#863 Lowering or suspending an account leaves its open live connections at their connect-time scope
  • ent#871 Audit the use of a self-change grant (capability_used) and unify grant audit names
  • ent#873 security: on hosted installs every sign-in shares one rate-limit bucket — 30 wrong logins from anyone lock everyone out for 5 minutes
  • ent#875 security: execution reads and agent control verbs sit at the accessor tier — a chat-shared user reads every other person's transcripts, terminates their turns, stops the agent, changes its model, force-resets git, mints any schedule's webhook (CWE-863)
  • ent#925 Peer skill grants: enforce in-container so a peer can't reach a library skill by asking in prose
  • ent#926 security: the accessor tier is the default for ~100 agent routes, and an agent acts for whoever chats with it — flip the default, scope the agent's turn to its person
  • ent#949 Files path policy: record rulings on the ent#820 deferred paths and sweep the remaining RUNS_CODE / secret-class siblings
  • ent#962 Owner view of a person's memory history, with undo — the seat-decision readers rule for MEM-001 (split from #419)
  • ent#980 Projects: a steward agent proposes members and task owners, the creator approves once in the Inbox
  • ent#986 Gemini runtime: set privacy.usageStatisticsEnabled: false so agents don't send Google a persistent ID + host fingerprint by default
  • ent#1001 Agent .env can set ANTHROPIC_BASE_URL: unprotected endpoint variable re-points Claude Code's API traffic (and its token) at any host
  • ent#1002 Gated-skill hook reads run identity from the parent claude's environment: a nested CLI session is checked as whatever execution id it sets
  • ent#1003 Gemini CLI runtime: chat runs have no wall-clock timeout, task runs time out only when output arrives
  • ent#1043 A skill already assigned stays usable when its new gate fails to write
  • ent#1053 Asks addressed to the approver seat: rule whether the answer is fenced to the seat like role-id asks
  • ent#1062 bug: seat decisions and voice reply accept a finished run, the shape t#4121 closed on the memory door
  • ent#1063 bug: an agent's own key can address a schedule to another person and fire it, minting a live seat run
  • ent#1066 Hand a credential to an agent in the Workspace chat: masked field, stored, never in the thread
  • ent#1068 Unassign a held skill when its declared gate cannot be written (follow-up to ent#1043)
  • ent#1073 Binding or revoking an agent's card funding can delete its other credentials
  • ent#1089 Projects: an agent requests member and owner changes; the agent's primary approves, then it applies
  • ent#1093 Claude Code runtime: set CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 so agents don't send Anthropic file-content fingerprints by default
  • ent#1101 External MCP access is off by default on a fresh install and switched on from Settings; existing installs stay as they are
  • ent#1165 Projects: scope the steward read to the project, and enforce the task-owner limit server-side
  • ent#1168 bug(security): agent SSH ports published on 0.0.0.0 by default; ufw hardening guide gives false confidence
  • ent#1225 bug(access): a platform continuation turn should stand as its origin turn — wakes and ask-resumes with nobody on the row read no executions
  • ent#1235 Projects: a per-project mode (Track / Propose / Auto) decides whether the steward may start work, and the platform enforces it
  • ent#1253 Channel /login sends verification emails with no rate limit (Telegram, WhatsApp, Slack)
  • ent#1258 A suspended account still signs in through the Workspace door and runs agent turns (v1.0.5-rc2)

Refactors

  • #2720 refactor(frontend): useVoiceSession.js bypasses the single API client — bare axios with a hand-built auth header
  • #3785 refactor: split operator_queue_service.py — hotspot #2, highest complexity in the backend, no covering issue
  • #4028 refactor(redis): move SingleFlightLock's refresh/release callers onto the atomic methods
  • #4096 ci: the per-PR unit job takes 24 min — xdist runs 2 workers on a 4-vCPU runner and db_backend rebuilds the schema per test (36% of suite time)
  • #4099 ci: journey-smoke boots a stack on every PR although Tier 2 runs it on dev — gate the PR run to lane C paths (+ label), fail-open

Private tracker (trinity-enterprise):

  • ent#801 Positioning (ADR-0011): clean the runtime system prompts — platform instructions header, meta-prompt copy, system-agent seed, ask_trinity, instance custom prompt

Documentation

  • #3964 docs(faq): fleet-design and practice questions from a new-org onboarding call
  • #4133 Docs: fork over MCP, the audit claim, and the new agent-facing tools and manifest fields

Private tracker (trinity-enterprise):

  • ent#797 Positioning (ADR-0011): OSS README, repo About, first-run UI, installer, marketplace listings and CLI onto the operating-system line

Other fixes (no type label)

  • #3840 bug(sanitizer): non-secret config values (FIBERY_HOST, GITHUB_REPO, TRINITY_MCP_URL) are redacted everywhere they appear, breaking links in agent replies
  • #3952 S-01 flags pull-owned rows as in_sql_only during the pre-enqueue window
  • #4440 [BUG] gemini-3.1-flash-live-preview stops working on November 17, 2026

Delivered inside the enterprise module

19 further private-tracker items shipped as part of the private enterprise module (pin 982735f → a26e522, 52 commits). Per the open-core docs rule (trinity-enterprise#45), they are not enumerated here.

Shipped in part (issues stay open)

  • #4175 Remove the SQLite backend — every compose file now defaults to the bundled PostgreSQL and SQLite needs an explicit setting; SQLite itself is not removed.
  • ent#1087 Direct call: interrupt the agent mid-reply — stop by voice or by hand shipped (#4211).
  • ent#641 Autonomy dial — the Workspace dial and the instance-level control shipped.
  • ent#965 Nevermined card delegation — steps 1 and 2 (ent#966, ent#967) shipped.
  • ent#1137 Projects as a working environment (epic) — the children listed above shipped; the epic stays open.
  • ent#1117 Scanner backlog — the dependency and code fixes reconciled from main (#4291, #4294, #4314, #4315, #4400) are in this release.
  • ent#1094 Marketplace image updates in a local /release run — the build, boot-check and submit scripts shipped (#4234).

Other changes

  • 11 dependency bumps (dependabot) on dev, plus the ent#1117 security patches and the llms.txt refresh (#4287) reconciled from main (#4600)
  • User-docs syncs (#4596, #4613), code-health dashboards and CSO reports

Abilities Marketplace (ecosystem)

Updates to the abilities plugin marketplace during this release window (2026-10-20 → 2026-11-17). Most landed as direct commits, referenced by SHA:

  • agent-dev: 46eed9b agent-dev 1.24.0 — project-management rev 2: standard template 2.0, project-init 1.6.0 (kickoff kit, participation table), project-steward 1.11
  • agent-dev: 78fb898 orchestrate 1.21 — a model floor on dispatch for chat_with_agent and fan_out
  • create-agent: e5a28ea create-website 1.4 (create-agent 2.4.0) — headless mode and token-based GitHub/Vercel delivery
  • trinity: Abilityai/abilities#27 trinity 2.12.1 — the onboard .gitignore scaffold excludes .trinity/ runtime state
  • Marketplace-wide: 06679ae Trinity 1.0.0 truth sync, 9a86a4a Trinity 1.0.5-rc2 truth sync

No marketplace releases were tagged in the window.

⚠️ Behavior changes / upgrade notes

PostgreSQL is the default; moving off SQLite is recommended

Every compose file now defaults to a bundled PostgreSQL (#4175 in part). SQLite reached
end-of-support on 2026-09-01 and has not been removed: it still runs in 1.0.5. We recommend
moving an install whose data is in trinity.db to PostgreSQL with
SQLITE_TO_POSTGRES.md, before or after the upgrade.

To keep running on SQLite, set DATABASE_URL= (empty) and TRINITY_ALLOW_SQLITE=1 in
.env. Without those two lines, start.sh and the images stop at startup on an install
that has SQLite data and print the copy command, so set them before upgrading if you are
not ready to move.

Use trinity update

scripts/deploy/update.sh (also trinity update) is the single entry point for updating a
self-hosted install (#3007): it runs a preflight before anything is touched, verifies the
published URL answers and rolls back if it does not.

Rebuild the agent base image and recreate agents

The base image moves from Node.js 20 to 24 (#3886) and carries the agent-side fixes in this
release (peer skill grants in-container, runtime privacy defaults, the Gemini process group
and timeouts). Agents on the old image keep running without them; a platform alert names
which running agents are still on the old image (#4483).

Behaviour that changes by default

  • Agent SSH ports publish on 127.0.0.1 (ent#1168). Applied when an agent is created
    or recreated; an existing agent keeps its binding until then. Set AGENT_SSH_BIND_ADDRESS
    to a VPN interface address, or 0.0.0.0 for the old behaviour.
  • External MCP access is off on a fresh install (ent#1101). An existing install keeps
    publishing :8080 on every interface until MCP_BIND_ADDRESS is set. Off-box clients are
    refused until an admin turns External MCP access on under Settings → MCP Keys.
  • Chat-shared people no longer read other people's executions or use agent control
    verbs
    (ent#875, ent#926). The accessor tier is no longer the default for agent routes.
  • The default turn limit is 200, up from 50 (ent#471).
  • Agents wake on an answered ask and on a delegated run ending (ent#790, ent#1108,
    ent#1102). Caller wakes are capped at 120 per hour per agent (CALLER_WAKE_MAX_PER_HOUR).
  • Terms of use are shown once to every first-time person on any door (ent#1107). The
    shipped default is admin-editable.
  • Agent runtimes send less to their vendors. Gemini usage statistics are off and Claude
    Code runs with CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 (ent#986, ent#1093).
  • ANTHROPIC_BASE_URL in an agent's .env is protected (ent#1001).
  • Prod and hosted compose add an ingress network on 172.30.0.0/24 (ent#873). The
    backend takes the client address only from the bundled nginx and Cloudflare Tunnel there.
    If that range is in use on the host, set TRINITY_PROXY_PREFIX.
  • A seeded default skill source that no admin retargeted follows the bundled pin on
    upgrade (ent#529).
  • The agent Sharing tab is named Channels (ent#225), and Credentials is its own Settings
    pane (ent#951).
  • Re-login after the backend restarts and reconnect MCP clients. Keyed MCP clients
    now survive an MCP server restart (#4060).

Config keys whose behaviour changed

Every .env / docker-compose*.yml key whose meaning changed between v1.0.0 and v1.0.5:

  • DATABASE_URL (all compose files) — unset now means the bundled PostgreSQL everywhere; in 1.0.0 that was true of hosted compose only, and prod and dev fell back to SQLite. Set-but-empty selects SQLite, which now also needs TRINITY_ALLOW_SQLITE=1.
  • POSTGRES_PASSWORD — required by every compose file (none renders without it; start.sh generates it). Previously required by hosted compose only.
  • POSTGRES_DB, POSTGRES_USER — honoured by dev compose only; prod and hosted fix both at trinity.
  • DB_POOL_SIZE, DB_MAX_OVERFLOW — unchanged values, but they now apply to every install, since PostgreSQL is the default.
  • DO_NOT_TRACK — widened. Any value other than unset or 0 now also turns off telemetry sharing, the help widget's Bug / Feature / Feedback report, the Ask tab and the ask_trinity MCP tool, and removes both vendor hosts from the Content-Security-Policy. Previously it disabled the operator-intake POST only (#4499).
  • VOICE_MODEL — default moves models/gemini-3.1-flash-live-preview → models/gemini-3.8-live; the old model stops working on 2026-11-17 (#4440).
  • FRONTEND_URL — invite emails now take their sign-in link from this key only. Empty, the email carries no link; previously the link was derived from request headers and could point at the public chat host (ent#1255). Set it whenever PUBLIC_CHAT_URL is set.
  • CANARY_ENABLED — true / yes / on now enable it as 1 does; an unrecognised value leaves it off and logs a warning (#1988).
  • VITE_BUG_REPORTING_ENABLED — build-time only: it applies where the frontend is built on the machine. To switch reporting off on the published images use the run-time BUG_REPORTING_ENABLED=false (#4499).
  • ASK_TRINITY_ENDPOINT — pointing it at an unreachable URL is no longer the way to switch the tool off; ASK_TRINITY_ENABLED=false unregisters it (#4499).
  • INTERNAL_API_SECRET — also signs MCP session ids, so rotating it costs every connected MCP client one reconnect (#4060).
  • FRONTEND_PORT — unchanged meaning, now combined with FRONTEND_BIND_ADDRESS in prod and hosted compose.

New keys (no prior meaning): TRINITY_ALLOW_SQLITE, TRINITY_POSTGRES_VOLUME,
MCP_BIND_ADDRESS, MCP_EXTERNAL_ACCESS, AGENT_SSH_BIND_ADDRESS, FRONTEND_BIND_ADDRESS,
CADDY_PROXY_PORT, TRINITY_PROXY_PREFIX, TRUSTED_PROXIES, BUG_REPORTING_ENABLED,
ASK_TRINITY_ENABLED, ALLOW_PDF_UPLOADS, DIRECT_VOICE_ENABLED, GEMINI_STT_MODEL,
GEMINI_TTS_MODELS, CALLER_WAKE_MAX_PER_HOUR, NEVERMINED_DELEGATION_MAX_INFLIGHT,
TRINITY_MEMORY_EMBEDDING_MODEL, FASTEMBED_CACHE_PATH (the last two are image-internal).
system_settings keys were not swept mechanically this release.

Enterprise instances

Bump the enterprise submodule pin to the release's commit (a26e522).

Known limitations

  • SQLite is still supported behind TRINITY_ALLOW_SQLITE=1 and leaves Trinity in a later
    release; moving to PostgreSQL is recommended (#4175).
  • Recreate agents on the 1.0.5 base image for the agent-side fixes (see above).
  • An existing install keeps its MCP port and its agents' SSH ports open until
    MCP_BIND_ADDRESS is set and each agent is recreated.
  • Retrieval memory ships switched off. Turn it on under Settings → General.
  • Direct calls: stop by voice or by hand shipped; the rest of mid-reply interruption is
    still in progress (ent#1087).
  • Funding an agent with a card covers enrolling a card and creating a mandate; the
    combined Integrations pane follows (ent#965).
  • Signed agent identity between instances stays opt-in (AAUTH_ENABLED).
  • Pull dispatch is an opt-in pilot per agent; push dispatch remains the default.
  • No freeze plan or coherence review was run for this cycle; the payload was tested
    through rc1–rc4.

Contributors

@dolho, @AndriiPasternak31, @vybe, @trinity-ability, @obasilakis, @webmixgamer,
@ericdelorefice, @pcbeingused333, @francisjervis, @alex-furman, @mayuriphad, @pavshulin —
plus dependabot.