# Omnius Reliability and Long-Horizon Workboard

Updated: 2026-08-29

This file is the durable workboard for the current repair program.
It records scope, dependencies, evidence, and closure rules.

## Status rules

- `[x]` means that code and focused tests are complete.
- `[ ] ACTIVE` means that work is in progress.
- `[ ] VERIFY` means that code exists, but a required gate is open.
- `[ ] QUEUED` means that an active dependency must finish first.
- `[ ] EXTERNAL` means that completion needs a repository or service action.

## P0: Telegram reliability

- [x] Trace the failing Telegram request from intake to final route.
  - Evidence: The router aborts after two 30-second request windows.
  - Evidence: The broker can still have an active or queued request.
  - Evidence: The plain retry is skipped after the liveness abort.
  - Required result: Telegram must not cancel healthy queued inference.
- [x] Compare Telegram transport with `AgenticRunner` transport.
  - Compare endpoint selection.
  - Compare native Ollama transport.
  - Compare broker admission.
  - Compare request timeout behavior.
  - Compare stream inactivity behavior.
  - Compare unary fallback behavior.
  - Compare thinking suppression.
  - Compare context sizing.
  - Compare usage accounting.
- [x] Replace the 30-second router deadline.
  - Do not count broker queue time as model failure.
  - Use a progress-aware inactivity limit.
  - Keep one bounded whole-request safety limit.
  - Do not start a duplicate unary request after an abort.
  - Preserve user cancellation.
- [x] Prevent duplicate inference after stream failure.
  - Retry only for a transport or protocol incompatibility.
  - Do not retry an aborted or timed-out generation as unary inference.
  - Release each broker slot exactly once.
- [x] Make the host broker authoritative for capacity.
  - Use the public broker endpoint at `http://127.0.0.1:11434`.
  - Do not start a local Ollama pool for this endpoint.
  - Preserve `think: false` and `num_ctx` on broker requests.
  - Preserve host lane selection and model aliases.
- [ ] ACTIVE — Reduce router prompt and call amplification.
  - Measure the silent-reflection prompt.
  - Measure the attention-router prompt.
  - Remove duplicate context.
  - Keep required persona, relationship, and thread evidence.
  - Avoid repair calls when no model token was returned.
- [x] Add Telegram timeout and retry regression tests.
  - Test a 45-second first-token delay.
  - Test a queued broker request.
  - Test a slow active stream.
  - Test an actual stalled stream.
  - Test user cancellation.
  - Test a protocol error that permits unary fallback.
  - Test an abort that forbids unary fallback.
- [x] Run focused Telegram suites.
- [x] Build the CLI package.
- [x] Stop treating a clean admin-chat answer as an incomplete agentic task.
  - Root cause: the live model returned a valid visible reply on each turn, but
    `AgenticRunner` required a `task_complete` tool call on the conversational
    Telegram surface. It therefore repeated the generation three times and
    committed the run as incomplete after 55,193 tokens and zero tool calls.
  - Add an explicit `acceptTextOnlyCompletion` runner contract. Enable it only
    for admin-DM chat; action and mutation runs retain authoritative completion
    and evidence gates.
  - Retain the first completed stream answer before completion-boundary
    filtering so Telegram cannot discard the terminal conversational reply.
- [x] Separate the Telegram admin reply from internal run diagnostics.
  - Render only the natural answer, concise work summary, or artifact
    description on the Reply tab.
  - Do not prepend `Admin run`, echo the request, append a page marker, expose
    raw exceptions, or persist completion reports as assistant conversation.
  - Put lifecycle status and mutations in Activity, tool results in Tools,
    completion and verification reports in Evidence, and prompt/context data in
    Context.
  - Return the panel to Reply when a run finishes, while marking updated
    internal tabs for inspection.
  - Preserve and deliver a usable draft and generated artifacts when the runner
    misses `task_complete`; keep the incomplete report in Evidence.
- [x] Add Telegram reply-versus-evidence regressions.
  - Assert exact clean Reply rendering for completed and failed panels.
  - Assert that the literal `back at it?` incomplete-run failure delivers only
    its usable conversational draft and stores `INCOMPLETE` behind Evidence.
  - Assert that a one-turn admin chat delivery and persisted assistant history
    contain the exact model answer, not the panel report.
- [ ] QUEUED — Make automatic generated-artifact upload receipts authoritative.
  - Return delivered and failed artifact receipts from the automatic sender.
  - Put raw upload diagnostics in Tools or Evidence.
  - If an attachment cannot be delivered, replace any success claim with one
    concise user-facing blocker without exposing a local path.
- [x] Make Telegram intake completion depend on a delivery receipt.
  - Require `ok: true` and a positive Telegram `message_id`.
  - Try the existing HTML-to-plain fallback, then fail the run if both sends
    fail or neither returns a receipt.
  - Mark the intake consumed only after a confirmed visible delivery; otherwise
    mark it failed for replay and diagnosis.
- [ ] VERIFY — Run live Telegram route inference.
  - Endpoint: `http://127.0.0.1:11434`.
  - Model: `robit/ornith-1.5:9b`.
  - Required GPU: a broker-selected A100 80 GB GPU.
  - Forbidden GPU: the GT 1030.
  - Partial live evidence: Telegram intake accepted `back at it?`, created a task
    epoch, and completed broker-backed `/v1/chat/completions` and `/api/chat`
    requests without the former 30-second abort.
  - The earlier open-ledger result is now diagnosed as a conversational
    completion-policy mismatch, not a transport or model failure.
  - Direct live proof after the causal fix: the exact Ornith model returned
    `TELEGRAM_CHAT_TERMINAL_OK`; the runner completed in one turn with zero tool
    calls and 30 completion tokens while resident on an A100 80 GB GPU.
  - Open gate: receive and deliver one fresh real Telegram message through the
    rebuilt installed CLI. Mocked Bot API coverage is complete, but it does not
    replace an end-to-end Telegram delivery receipt.
  - Ready state: the rebuilt `/usr/local/bin/omnius` TUI is running in
    `telegram_test`; its Telegram poller owns the live workspace lock.
  - Required result: valid visible route JSON.
  - Required result: nonzero generated token telemetry.
  - Required result: no 30-second abort.

## P0: Router empty-content failure

- [x] Repair the native Ollama structured-output regression exposed by live
  Telegram messages 4594, 4597, and 4598.
  - Active-build evidence: version 1.0.653 completed both router generations,
    but each returned an out-of-domain `{tool,args}` object. Intake 4597 was
    then consumed as `shouldReply:false` before an action runner was created.
  - Root cause: both preferred native Ollama transports reduced every OpenAI
    `json_schema` response format to `format:"json"`. The router's required
    fields and `additionalProperties:false` constraint never reached Ollama.
  - Preserve `json_object` as native `"json"`, but pass the complete inner JSON
    Schema object to native unary and streaming `/api/chat` requests.
  - Preserve typed native tool calls across unary and Telegram stream
    accumulation. Do not interpret structured-output text as a tool call.
  - Keep wrong-domain public router output quarantined. Router-supplied tool
    arguments do not execute and do not bypass public attention policy.
  - Add an explicit router classification boundary: capability needs become
    `route:"action"`; the attention router does not invoke tools.
- [x] Add native-schema and action-admission regressions.
  - Orchestrator native transport suite: 22 passed.
  - Telegram Bot API suite: 103 passed, including the new public action
    admission path from router schema to AgenticRunner and delivered reply.
  - Telegram observability and stream-lifecycle suites: 47 passed.
  - Orchestrator and CLI TypeScript builds passed.
- [ ] VERIFY — After operator activation, repeat the public-group action request
  and confirm the first router response is a schema-conforming
  `{route:"action",should_reply:true,...}` decision followed by one action run
  and a Telegram delivery receipt.
- [x] Separate Ollama transport detection from local pool ownership.
- [x] Use native `/api/chat` for broker-managed Ollama control requests.
- [x] Send `think: false` through the broker transport.
- [x] Send `num_ctx` through the broker transport.
- [x] Strip prior hidden-thinking blocks from model history.
- [x] Record JSON-mode and plain-mode outcomes separately.
- [x] Preserve exact backend errors in router diagnostics.
- [x] Skip repair when the backend returned no visible text.
- [x] Remove the premature Telegram liveness abort.
- [ ] VERIFY — Confirm that the route parser receives visible JSON live.

## P0: Model broker and parallel inference

- [x] Confirm that the host broker is available.
- [x] Confirm that the parallel pool is enabled.
- [x] Confirm a maximum of three managed servers.
- [x] Confirm one in-flight request per managed lane.
- [x] Confirm that three A100 80 GB GPUs are selected.
- [x] Confirm that the GT 1030 is not selected.
- [x] Confirm that the 9B model is resident on an A100.
- [x] Confirm that no Ollama runner is resident on the GT 1030.
- [x] Confirm the exact model tag `robit/ornith-1.5:9b`.
- [x] Confirm model context length `262144`.
- [x] Confirm model capabilities for completion, vision, tools, and thinking.
- [x] Replace per-request capacity polling with one FIFO scheduler and one reconciler.
- [x] Cancel disconnected queued clients before backend inference starts.
- [x] Add queue IDs, positions, phases, depth, age, and aggregate timing metrics.
- [x] Make queue time exclude backend generation and time to first response.
- [x] Warm each managed model with zero generated tokens before lane readiness.
- [x] Verify model residency before admitting a lane.
- [x] Canonicalize only an omitted `:latest` tag.
- [x] Preserve separately configured model aliases.
- [x] Replace least-recently-used idle lanes for mixed-model demand.
- [x] Remove the system-lane block on independent managed-lane expansion.
- [x] Scope anonymous CUDA handling to selected GPUs and broker-external processes.
- [x] Ignore foreign-process departure and unselected-GPU churn.
- [x] Retire only affected managed lanes after stable allocation growth.
- [x] Protect queued timeout and client-write paths from broken-pipe tracebacks.
- [x] Keep `/api/tags`, `/api/ps`, `/api/version`, `/api/show`, and `/v1/models`
  available during cooperative GPU-lease drains.
- [x] Add broker regressions for lazy three-lane scaling, FIFO, cancellation, aliases,
  mixed models, warm readiness, queue telemetry, and CUDA process churn.
- [x] Add the negotiator suite to the broker repository CI workflow.
- [x] Commit broker scheduler repair locally on `main` as `4fbad0c`.
- [x] Commit drain-safe metadata repair locally on `main` as `cd344e3`.
- [x] Commit host-memory restart preflight repair on `main` as `4a24eb5`.
- [x] Commit endpoint-aware queue admission repair on `main` as `4349a5b`.
- [x] Commit the public private-lane range contract on `main` as `c98492f`.
- [x] Deploy the repaired broker through `./ollama-unify.sh --install-safety`.
  - Installed helper SHA-256:
    `39458260611a80f50d87f03e7567bcdcc632ebd1af421e9d94f69521410d80cd`.
  - `ollama-unify-negotiator.service` and `ollama.service` are active.
  - Public port `11434` and system backend port `11436` report Ollama `0.32.13`.
- [x] Disable the conflicting legacy `ollama-env-update.service` unit.
- [x] Confirm that Omnius can request broker capacity without bypass.
  - Discovery published managed ports `11437` through `11468`.
  - Omnius created no private `ollama serve` owner during the live run.
- [x] Confirm parallel requests on separate host lanes.
  - Live discovery showed two concurrent `robit/ornith-1.5:9b` lanes on separate
    A100 80 GB GPUs and a separate embedding lane.
- [ ] VERIFY — Confirm that broker queue time remains visible to Telegram.
- [x] Confirm broker commits `4fbad0c`, `cd344e3`, `4a24eb5`, `4349a5b`, and
  `c98492f` are pushed to `origin/main`.
- [x] Confirm local and remote broker `main` are identical at `c98492f`.
- [x] Confirm the required broker service restart state.

## P0: Broker ownership incident root cause

- [x] Identify the duplicate scheduler and port-ownership race.
  - Root cause: an older Omnius elastic-pool worker survived independently and
    spawned detached Ollama servers on ports `11436` through `11438` while the
    broker installer temporarily stopped the supervised services.
  - Result: the system Ollama service could not reclaim `11436`, and the host
    broker opened another managed set on later ports.
- [x] Remove the stale Omnius worker and its detached Ollama process groups.
- [x] Route broker-managed setup and recovery through systemd.
- [x] Fail closed when local broker discovery is malformed or temporarily absent.
- [x] Remove the unsafe `OMNIUS_OLLAMA_UNIFY_ALLOW_POOL` ownership bypass.
- [x] Prevent `/parallel` from detached-spawning Ollama for a broker-managed endpoint.
- [x] Keep one supervised system Ollama owner and one broker scheduler.

## P0: Runtime package and daemon integrity

- [x] Replace the Zod-4-only `z.json()` runtime contract with a recursive JSON
  schema that works with the bundled Zod runtime.
- [x] Add nested JSON acceptance and `undefined` rejection tests.
- [x] Make `publish/dist/launcher.cjs` executable during every publish build.
- [x] Confirm the source artifact and `/usr/local` installed CLI have the same
  SHA-256 and report Omnius `1.0.649`.
  - Current installed `dist/index.js` SHA-256:
    `9e7a830d1370c0127f38fa4e24b8b15cc9069746904d80a26e41f9ee48a59b8b`.
- [x] Remove the stale user-service drop-in that moved the shared daemon from
  port `11435` to the Cygnus port `11535`.
- [x] Confirm the shared Omnius daemon is active on `11435` and the separate
  Cygnus daemon is active on `11535`, both at `1.0.649`.
- [x] Replace the stale `telegram_test` main-model identifier with the exact tag
  `robit/ornith-1.5:9b`.
- [x] Stop the diagnostic TUI and remove its dead-PID Telegram ownership lock.

## P0: Startup control-plane recovery

- [x] Replace boolean model availability with `available`, `unavailable`, and
  `unknown` outcomes.
- [x] Treat timeout, HTTP 502, and connection failure as `unknown`.
- [x] Preserve the configured model when availability is unknown.
- [x] Run first-run setup only after a successful catalog proves the model absent.
- [x] Match omitted Ollama `:latest` with strict canonical equality.
- [x] Use one coalesced names-only `/api/tags` request for startup availability
  and readline completion.
- [x] Keep full per-model `/api/show` discovery out of startup checks.
- [x] Classify Telegram startup failures as retryable or terminal.
- [x] Retry network errors, HTTP/Bot API 429, 5xx, and webhook propagation delay.
- [x] Do not retry invalid tokens or duplicate long-poller ownership conflicts.
- [x] Add a bounded TUI Telegram auto-start supervisor.
- [x] Cancel pending Telegram auto-start when the operator stops Telegram.
- [x] Use one daemon `/health` response for liveness and immutable boot identity.
- [x] Tolerate bounded transient daemon health-read failures before recovery.
- [x] Require daemon health payload `status: ok` before accepting identity.

## P0: Live inference hardware gate

- [x] Read `/usr/local/share/ollama-unify/AGENTS.md`.
- [x] Run `docker gpu discover`.
- [x] Confirm public endpoint `http://127.0.0.1:11434`.
- [x] Confirm backend endpoint `http://127.0.0.1:11436`.
- [x] Inspect `/api/ps`.
- [x] Inspect `/api/tags` for the exact model.
- [x] Inspect `/api/show` for the exact model.
- [x] Inspect `nvidia-smi` GPU inventory.
- [x] Inspect `nvidia-smi` model process placement.
- [x] Confirm model runner PID `2199237` on an A100.
- [x] Confirm that no test runner uses the GT 1030.
- [x] Send the first generated-token request.
  - Exact response text: `BROKER_OK`.
  - Provider usage: 5 evaluated tokens.
- [x] Inspect GPU placement during generation.
  - The Ornith runner was resident on an A100 80 GB GPU.
  - No Ollama runner was resident on the GT 1030.
- [x] Re-preflight the final installed live session.
  - The broker admitted zero-token capacity on managed lane 10.
  - Runner PID `3179075` is resident on A100 GPU 2.
  - No Ollama runner is resident on the GT 1030.
- [x] Record response text, provider usage, and broker admission.
- [ ] VERIFY — Record live ontology graph validity.

## P1: Ontology long-horizon work graph

- [x] Add a versioned operational-world schema.
- [x] Add strict work-node schemas.
- [x] Add strict relation schemas.
- [x] Add source-binding schemas.
- [x] Add evidence-reference schemas.
- [x] Add structured claim schemas.
- [x] Add first-class conflict schemas.
- [x] Add action-contract schemas.
- [x] Add world-patch schemas.
- [x] Add validation-result schemas.
- [x] Add immutable validation-receipt schemas.
- [x] Add bounded reasoning-slice schemas.
- [x] Export all schemas from `@omnius/schemas`.
- [x] Add a deterministic world reducer.
- [x] Reject stale base revisions.
- [x] Reject reused target revisions.
- [x] Reject duplicate patch targets.
- [x] Reject dangling references.
- [x] Reject duplicate semantic relations.
- [x] Reject invalid decomposition cycles.
- [x] Reject invalid temporal intervals.
- [x] Require evidence for verified relations and claims.
- [x] Preserve incompatible claims as conflicts.
- [x] Cascade node removal when a caller requests it.
- [x] Preserve full graph snapshots in revision history.
- [x] Emit deterministic graph and evidence hashes.
- [x] Persist operational-world state atomically.
- [x] Restore valid state at a task boundary.
- [x] Reject invalid persisted state.
- [x] Project the active objective.
- [x] Project todo hierarchy.
- [x] Project requirements.
- [x] Project acceptance criteria.
- [x] Project features, components, modules, and tasks.
- [x] Project source documents and sections.
- [x] Project target paths, symbols, documents, and external systems.
- [x] Project action nodes.
- [x] Project mutation nodes.
- [x] Project blocker nodes.
- [x] Project completion claims.
- [x] Project completion evidence.
- [x] Project trusted command receipts.
- [x] Project durable workboard cards, dependencies, assignees, decisions,
  blockers, source bindings, verifier contracts, and evidence.
- [x] Keep the workboard authoritative for execution while the graph remains
  its reasoning and proof projection.
- [x] Link explicit completion claims to the matching todo proof path.
- [x] Require task-to-verification-to-evidence closure for explicit work.
- [x] Reject a todo status as proof for acceptance, artifacts, or verification.
- [x] Reject failed, timed-out, stale, malformed, and model-forged verifier
  receipts.
- [x] Keep trusted command receipts out of model-callable workboard schemas.
- [x] Preserve authority as explicit, adopted, or inferred.
- [x] Add the constrained controller boundary for model-proposed graph patches.
- [x] Reject model authority escalation.
- [x] Build target-focused graph slices.
- [x] Include supporting evidence.
- [x] Include refuting evidence.
- [x] Include open conflicts.
- [x] Preserve parent reference closure.
- [x] Preserve action-target closure.
- [x] Limit model-visible graph context.
- [x] Enforce a whole rendered-context character budget.
- [x] Report partial coverage and omitted-node counts.
- [x] Render relations only when both endpoints are visible.
- [x] Refresh the graph before each primary model turn.
- [x] Refresh the graph before each brute-force model turn.
- [x] Keep one current graph block in context.
- [x] Gate only explicit graph obligations.
- [x] Keep inferred graph content advisory.
- [x] Recheck graph authority on all terminal paths.
- [x] Document the implementation in `docs/ONTOLOGY-LONG-HORIZON-WORK-GRAPH.md`.
- [ ] VERIFY — Run live 9B graph decomposition inference.
- [ ] VERIFY — Parse the model patch with `WorldPatchSchema`.
- [ ] VERIFY — Apply the model patch with the deterministic reducer.
- [ ] VERIFY — Record a successful validation receipt.

## P1: Todo ontology contract

- [x] Add `workKind`.
- [x] Add `requirementIds`.
- [x] Add `acceptanceIds`.
- [x] Add `sourceRefs`.
- [x] Add `targetRefs`.
- [x] Add typed action contracts.
- [x] Preserve all fields in the todo store.
- [x] Preserve all fields in todo tool input parsing.
- [x] Preserve all fields in no-change detection.
- [x] Export all new types.
- [x] Add persistence and update tests.

## P1: Feature and completion authority

- [x] Invoke `verifyUnit` after accepted execution.
- [x] Fail the feature unit when verification returns false.
- [x] Fail the feature unit when verification throws.
- [x] Contain and persist verifier errors.
- [x] Prevent fallback completion without evidence.
- [x] Preserve feature-node graph, card, requirement, acceptance, evidence,
  verification-receipt, and attempt lineage.
- [x] Save feature trees with atomic replacement.
- [x] Distinguish absent delivery manifests from invalid manifests.
- [x] Reject invalid authored delivery manifests.
- [x] Reject incomplete authored delivery manifests.
- [x] Require fresh verifier evidence after relevant mutations.
- [x] Keep frontier and inferred claim feedback advisory.
- [x] Add a bounded completion-hold escape.
- [x] Return `incomplete_verification` after repeated authoritative holds.
- [x] Apply the same fail-safe to alternate terminal paths.

## P1: Telegram temporal relationship flow

- [x] Add immutable scoped turn nodes.
- [x] Record direct replies, mentions, participation, and turn order as typed
  relations.
- [x] Separate observed, reported, inferred, and operational evidence classes.
- [x] Keep router relationship notes as advisory claims.
- [x] Build one-hop target-focused social context for the current chat and
  thread.
- [x] Preserve relation observations and temporal validity instead of only a
  lossy aggregate.
- [x] Stop ordinary group-chat text from entering global user personality
  memory.
- [x] Persist Telegram conversation snapshots atomically.
- [x] Add an append-only canonical Telegram social event log.
- [x] Replay missing social events into the snapshot after an interrupted
  observation-to-snapshot boundary.
- [x] Consume Telegram `message_reaction` updates.
- [x] Consume Telegram `message_reaction_count` updates.
- [x] Track the actor, target message, target participant, added reactions,
  removed reactions, and current reaction set.
- [x] Record reactions sent by Omnius through `setMessageReaction`.
- [x] Keep emoji meaning unclassified in the transport graph. Any feeling or
  intent interpretation stays advisory and must cite the observed reaction.
- [ ] QUEUED — Add deterministic alias-merge receipts.
- [ ] QUEUED — Add a multi-store social projection transaction coordinator if
  the durability requirement expands beyond atomic per-store writes.

## P1: Daemon frontend redesign

- [x] Capture the bounded Huggle desktop, collapsed-sidebar, menus, Explorer,
  and mobile interaction references.
- [x] Map the current `localhost:11435` frontend and its regression suites.
- [x] Rebuild the daemon shell, responsive navigation, chat composer,
  capability cards, model picker, attachment menu, and Explorer interaction
  hierarchy from the bounded reference.
- [x] Preserve Omnius identity, owned assets, accessibility, daemon
  routes, diagnostics, and operational capabilities.
- [x] Update route, script, layout, responsive, and interaction tests.
- [x] Remove generic keyword-class coloring for `error`, `success`, `ok`,
  `bad`, and `live`; explicit state components retain their own presentation.
- [x] Fix native `/v1/chat` streaming through an active remote share.
  - Detect `{stream:true}` from the request contract instead of relying on an
    `Accept` header that the native chat client does not send.
  - Forward remote HTTP status, safe response headers, and SSE bytes without a
    nested SSE wrapper.
  - Preserve the original abort signal. User cancellation does not clear the
    selected remote share.
  - Clear a rejected sticky remote-share transport without automatically
    resubmitting a possibly accepted chat mutation to the local daemon.
  - Accept Tor-only remote-share state.
- [x] Decouple daemon-owned chat work from the browser response lifetime.
  - Guard all child-callback SSE writes and final response closure after a tab
    closes, reloads, or loses its proxy transport.
  - Keep the child running and persist its terminal in-flight result for
    reattachment.
  - Handle child spawn errors without an unhandled process event.
  - Mark the server-side chat registry from child completion or failure, not
    from browser connection closure.
  - Verify a pre-inference spawn failure returns accepted, structured failure,
    and DONE frames while `/health` remains available.
- [x] Keep the operational-world JSON schema compatible with the Zod runtime
  embedded in the publish bundle. Do not use the Zod 4-only `z.json()` helper
  on the daemon startup path.
- [ ] QUEUED — Replace parent-owned child stdout/stderr pipes with a durable
  supervised output journal before claiming that detached chat children
  survive a daemon restart. A detached PID alone does not make its pipes
  restart-safe.
- [x] Verify the rebuilt frontend on desktop and mobile viewports.
- [x] Keep publish, install, daemon activation, and live service replacement
  outside this source-only work scope. The operator owns those actions.

## P1: Public Telegram presentation leak

- [x] Prevent `[STEERING_RECONCILIATION]` contracts from reaching a
  public Telegram reply.
- [x] Prevent internal `no_reply` rationale, identity-boundary notes,
  and injected-chatter analysis from reaching a public Telegram reply.
- [x] Keep these records available only in the internal Context or
  Evidence tabs.
- [x] Add public-chat regressions for addressed-to-other-bot,
  no-reply, steering reconciliation, and ordinary group chatter cases.

## P1: Latest Hermes session and durable-work audit

- [x] Audit official `NousResearch/hermes-agent` at commit
  `1a47a36422ae3c42c23839b72cc24fb6a651aa28` (2026-08-29).
- [x] Compare session lineage, prompt snapshots, strict role alternation,
  memory-provider lifecycle, turn ownership, delivery obligations, durable
  Kanban claims, and stale-claim recovery with Omnius.
- [x] Preserve Omnius strengths that Hermes does not replace: typed work graph,
  claim/evidence separation, fresh verifier receipts, mutation lineage, and
  fail-closed completion projection.
- [x] Adopt the relevant event-sourcing pattern for Telegram social state.
- [x] Keep system-prompt snapshots stable within one active conversation.
- [x] Keep platform identity and capability scope in session-owned state, not
  process-global environment state.
- [ ] QUEUED — Add a durable outbound Telegram delivery-obligation ledger with
  idempotent message keys and crash recovery.
- [ ] QUEUED — Bind active long-horizon work claims to renewable ownership
  leases and reclaim stale claims without discarding the task history.
- [ ] QUEUED — Add session-lineage edges for compression, reset, delegation,
  and recovery boundaries to the operational world graph.
- [ ] QUEUED — Add deterministic shutdown receipts for background memory writes
  and report abandoned writes separately from abandoned prefetch work.

## P1: Inference visibility

- [x] Add a shell-style inference block.
- [x] Show elapsed inference time.
- [x] Show request kind, model, and session.
- [x] Show handling state before the first token.
- [x] Show thinking-token regions to the operator.
- [x] Show visible content-token regions to the operator.
- [x] Show streamed tool argument regions.
- [x] Show unary returned contracts before parsing.
- [x] Read usage from structured stream chunks.
- [x] Show an estimated generated-token count.
- [x] Add a read-mode control.
- [x] Expand all prior generated rows in read mode.
- [x] Keep thinking content out of Telegram replies.
- [x] Show broker queue admission as a separate phase.
- [x] Show time to first byte separately from queue time.
- [x] Stop showing successful `~0 tok` when provider usage is available.

## P1: Clickable files, folders, and websites

- [x] Add one shared terminal-link utility.
- [x] Convert Markdown links to OSC 8 links.
- [x] Convert bare `http` and `https` URLs to OSC 8 links.
- [x] Convert absolute file paths to `file://` links.
- [x] Convert existing relative project paths to absolute `file://` links.
- [x] Remove line and column suffixes from the open target.
- [x] Keep line and column suffixes in visible text.
- [x] Link directories so the desktop file manager opens them.
- [x] Link files so the default file handler opens them.
- [x] Link websites so the default browser opens them.
- [x] Apply links to completed assistant Markdown.
- [x] Apply links to streamed assistant lines.
- [x] Apply links to Telegram diagnostics and sub-agent output.
- [x] Apply links to tool Markdown output that enters natural-language rendering.
- [x] Do not alter code blocks, commands, JSON, logs, or machine payloads.
- [x] Add terminal-link unit tests.
- [x] Add streaming-link regression tests.
- [x] Add Telegram-link regression tests.

## P1: ASD-STE100 communication policy

- [x] Review ASD-STE100 Issue 9.
- [x] Add one shared STE communication policy.
- [x] Apply the policy to normal agent tiers.
- [x] Apply the policy to router prompts.
- [x] Apply the policy to Nexus requests.
- [x] Preserve code, JSON, commands, paths, identifiers, and quotations.
- [x] Preserve logs, tool output, and required machine formats.
- [x] Add prompt policy tests.
- [x] Update stale Nexus prompt tests.
- [ ] VERIFY — Review remaining dynamically generated tool instructions.
- [ ] VERIFY — Review remaining Telegram public prompts.
- [ ] VERIFY — Run a qualified human conformance review before certification.

## P2: 2026 reliability practices

- [x] Use bounded intervention replay checkpoints.
- [x] Use evidence-grounded completion.
- [x] Use fresh post-mutation verification.
- [x] Use explicit authority boundaries.
- [x] Keep inferred scope advisory.
- [x] Use target-focused context slices.
- [x] Preserve contradictions instead of overwriting them.
- [x] Use deterministic reducers for model-proposed state.
- [x] Record immutable validation receipts.
- [x] Use bounded repeated-hold escape behavior.
- [ ] VERIFY — Recheck current 2026 primary literature before the final audit.
- [ ] VERIFY — Add exact citations to the final engineering review.

## Verification record

- [x] Schema tests: 230 passed.
- [x] Todo-store tests: 39 passed.
- [x] Operational reducer tests: 9 passed.
- [x] Operational controller tests: 6 passed.
- [x] Operational runner test: 1 passed.
- [x] Feature-node tests: 7 passed.
- [x] Workboard trusted-receipt tests: 22 passed.
- [x] Operational work-graph, reducer, feature-node, and context replay tests:
  30 passed.
- [x] Telegram temporal relationship regression, voice-soul, and Bot API tests:
  103 passed.
- [x] Daemon frontend route, script, responsive-layout, and interaction
  regressions: 15 passed.
- [x] Headless Chromium verified desktop and mobile live layouts, capability
  counts, menu behavior, empty-state transitions, route titles, and mobile
  drawer behavior with zero page errors.
- [x] Delivery coverage focused tests passed.
- [x] Workspace TypeScript builds passed.
- [x] Targeted Nexus tests passed.
- [x] Two full-suite timeout cases passed when run without worker contention.
- [x] Telegram transport, observability, link, and Bot API suites: 161 passed.
- [x] Native Ollama transport suite: 19 passed.
- [x] Model-broker focused suite: 6 passed.
- [x] Model availability and first-run setup regressions: 11 passed.
- [x] Telegram Bot API, startup retry, and routing suite: 91 passed.
- [x] Daemon identity, singleton, and listener-reclaim regressions: 23 passed.
- [x] CLI typecheck passed after startup control-plane repairs.
- [x] CLI build passed after startup control-plane repairs.
- [x] Inference visibility, terminal links, Telegram transport, and GPU policy:
  81 focused tests passed.
- [x] Completion authority, evidence, feature verification, and broker ownership:
  79 focused orchestrator tests passed.
- [x] `task_complete`, broker execution, and todo persistence: 46 focused tests passed.
- [x] Cross-version operational-world JSON schema: 15 tests passed.
- [x] Published and installed launchers are mode `0755` and their bundled CLI
  hashes match.
- [x] Rerun Telegram focused suites after the timeout repair.
- [x] Telegram routing, inference, observability, and receipt regressions after
  the conversational completion repair: 158 passed.
- [x] Functional admin-chat regression proves one tool-bearing runner inference,
  exact visible reply delivery, positive message receipt, and one persisted
  assistant history record.
- [x] Telegram Reply/Evidence isolation suite: 101 focused tests passed, including
  the literal incomplete `back at it?` delivery regression.
- [x] CLI typecheck, all workspace builds, publish generation, and publish
  artifact audit passed after Reply/Evidence isolation.
- Operator-owned activation note: rebuilding, installing, restarting, and
  activating the daemon or `telegram_test` poller are not tasks on this board.
- [x] Direct live `AgenticRunner` inference with `robit/ornith-1.5:9b` completed
  in one turn with 5,837 total provider tokens and no `task_complete` demand.
- [x] Legacy `maxTurns` control passed alone with a 60-second harness timeout
  (20.9-second test body); its earlier 30-second failure occurred only under
  concurrent test-worker contention.
- [x] Rerun CLI typecheck and build after link support.
- [x] Rerun all workspace TypeScript builds after the transport repair.
- [x] Rebuild all workspaces, regenerate `publish/`, pass the publish artifact
  audit, install the rebuilt package, and restart both user daemons.
- [ ] VERIFY — After operator activation, observe one live Telegram route and
  one live ontology patch without changing service ownership from this task.

## Closure rule

Do not close this workboard after a code-only result.

Close it only when these conditions are true:

1. Telegram returns a valid route through the selected broker model.
2. No healthy broker request fails at 30 seconds.
3. Generated contract tokens are visible in the inference block.
4. The operator can expand the complete inference transcript.
5. Files, folders, and websites are clickable in natural-language output.
6. The 9B model produces a valid ontology patch.
7. The reducer accepts the patch and emits a receipt.
8. Focused tests and workspace builds pass.
