omnius 1.0.689 → 1.0.691

package/docs/DISCOVERY.json

@@ -33824,6 +33824,84 @@
         }
       ]
     },
+    {
+      "id": "guide.work-orders-runtime-health-remediation-wo-21-task-convergence-and-todo-scope-uppercase",
+      "kind": "guide",
+      "title": "WO-21: Task convergence and todo scope",
+      "summary": "On 2026-09-03, a user reported that Omnius remained at task 2/8 for about 40 turns. The report did not include a screenshot or run artifact, so the exact label remains unverified. Omnius has two similar counters:",
+      "keywords": [
+        "work",
+        "orders",
+        "runtime",
+        "health",
+        "remediation",
+        "WO",
+        "21",
+        "task",
+        "convergence",
+        "and",
+        "todo",
+        "scope",
+        "md"
+      ],
+      "maturity": "internal",
+      "audiences": [
+        "maintainer",
+        "large-context-agent"
+      ],
+      "layer": "documentation",
+      "interfaces": [
+        {
+          "type": "file",
+          "target": "docs/work-orders/runtime-health-remediation/WO-21-task-convergence-and-todo-scope.md"
+        }
+      ],
+      "references": [
+        {
+          "type": "documentation",
+          "target": "docs/work-orders/runtime-health-remediation/WO-21-task-convergence-and-todo-scope.md",
+          "relation": "canonical-artifact"
+        }
+      ]
+    },
+    {
+      "id": "guide.work-orders-telegram-dmn-wo-22-dmn-outreach-and-learning-uppercase",
+      "kind": "guide",
+      "title": "WO-22: DMN outreach, DM sharing, and outcome learning",
+      "summary": "On 2026-09-03 at 17:07 PDT the bot posted in the OMNIUS group without being addressed. The operator asked whether this was self-induced reflection.",
+      "keywords": [
+        "work",
+        "orders",
+        "telegram",
+        "dmn",
+        "WO",
+        "22",
+        "dmn",
+        "outreach",
+        "and",
+        "learning",
+        "md"
+      ],
+      "maturity": "internal",
+      "audiences": [
+        "maintainer",
+        "large-context-agent"
+      ],
+      "layer": "documentation",
+      "interfaces": [
+        {
+          "type": "file",
+          "target": "docs/work-orders/telegram-dmn/WO-22-dmn-outreach-and-learning.md"
+        }
+      ],
+      "references": [
+        {
+          "type": "documentation",
+          "target": "docs/work-orders/telegram-dmn/WO-22-dmn-outreach-and-learning.md",
+          "relation": "canonical-artifact"
+        }
+      ]
+    },
     {
       "id": "guide.work-orders-telegram-dropbear-context-rca-workorder",
       "kind": "guide",

package/docs/DISCOVERY.md

@@ -559,6 +559,8 @@ Daemon equivalents are `GET /v1/discovery/bootstrap`, `GET /v1/discovery?q=<inte
 | `guide.work-orders-runtime-health-remediation-wo-19-clean-build-evidence-uppercase` | WO-19 Clean Build Evidence | WO-19 is implemented in commit 5f253c05. |
 | `guide.work-orders-runtime-health-remediation-wo-19-clean-build-reproducibility-uppercase` | WO-19: Clean Build and Optional-Dependency Reproducibility | Status: deterministic acceptance complete Risk: medium Depends on: WO-00 |
 | `guide.work-orders-runtime-health-remediation-wo-20-exact-session-continuation-uppercase` | WO-20: Exact session continuation across exit and restart | On 2026-09-03, the telegramtest TUI displayed the latest pentest task during startup, but the first restored model context combined a completed breach.py handoff with an unrelated, older page.tsx todo tree. The user then entered continue, and Omnius continued the stale tree instead of the task that was active before /quit. |
+| `guide.work-orders-runtime-health-remediation-wo-21-task-convergence-and-todo-scope-uppercase` | WO-21: Task convergence and todo scope | On 2026-09-03, a user reported that Omnius remained at task 2/8 for about 40 turns. The report did not include a screenshot or run artifact, so the exact label remains unverified. Omnius has two similar counters: |
+| `guide.work-orders-telegram-dmn-wo-22-dmn-outreach-and-learning-uppercase` | WO-22: DMN outreach, DM sharing, and outcome learning | On 2026-09-03 at 17:07 PDT the bot posted in the OMNIUS group without being addressed. The operator asked whether this was self-induced reflection. |
 | `guide.work-orders-telegram-dropbear-context-rca-workorder` | Telegram Dropbear Context Engineering RCA Work Order | Observed run: /home/roko/Documents/Projects/Adjacent/telegramtest/.omnius, run id 1782873796963-i5r7mv. |
 | `guide.work-orders-wo-am-gaps-uppercase` | Associative Memory Gap Work Orders | Generated: 2026-04-13 Source: Deep audit of multimodal associative memory systems Status: READY FOR IMPLEMENTATION |
 | `guide.work-orders-world-class-memory-compiler-readme-uppercase` | World-Class Memory Compiler Program | Status: active implementation program Owner: Omnius orchestration and memory packages Last updated: 2026-07-13 |

package/docs/work-orders/runtime-health-remediation/TRACKER.md

@@ -311,3 +311,15 @@ deterministically repaired P0 or P1 defect.
   pass cancellation, restart, ownership, external-effect, and isolation tests.
 - [x] WO-20 exit and restart preserve one exact typed task continuation. Generic
   context restoration cannot mix or activate unrelated historical artifacts.
+- [x] WO-21 task convergence and todo scope prevent stale `tasks N/M` displays,
+  cross-runner todo leakage, and activity-only re-engagement.
+  - [x] Bind todo tools to immutable runner-owned sessions.
+  - [x] Clear the active persisted checklist at a fresh task boundary after
+    archival.
+  - [x] Count only typed authoritative advancement as progress.
+  - [x] Latch an advisory convergence review when the leaf frontier is static.
+  - [x] Make loop, reminder, budget, and workboard progress semantics truthful.
+  - [x] Pass deterministic race, stasis, display, typecheck, and build tests.
+- [x] A recovered pause no longer strands its session. Submitting a new task
+  retires the superseded generation instead of refusing every later prompt,
+  while unreconciled external effects still fence admission.

package/docs/work-orders/runtime-health-remediation/WO-21-task-convergence-and-todo-scope.md

@@ -0,0 +1,115 @@
+# WO-21: Task convergence and todo scope
+
+## Incident
+
+On 2026-09-03, a user reported that Omnius remained at `task 2/8` for about
+40 turns. The report did not include a screenshot or run artifact, so the
+exact label remains unverified. Omnius has two similar counters:
+
+- `tasks 2/8` is the completed-leaf count in the pinned TUI checklist.
+- `Loop intervention 2/8` is the second repetition intervention in a
+  model-tier-specific series.
+
+The source audit found defects in both paths. These defects are sufficient to
+produce the reported symptom even though they do not prove which counter the
+reporter saw.
+
+## Root causes
+
+- A TUI process reuses one todo session across tasks. A fresh runner hides old
+  todo IDs in its private view but leaves the persisted checklist unchanged.
+  The TUI reads that unchanged file and can display a stale `tasks 2/8` row.
+- Todo tools fall back to one mutable process-global session ID. Concurrent
+  parent, Telegram, background, and child runners can redirect one another's
+  unscoped reads and writes.
+- Any unique non-noop tool result increments the counter used to permit turn
+  extension and brute-force re-engagement. Varying reads, searches, and error
+  text therefore masquerade as task advancement.
+- The repeated-loop counter resets after its maximum intervention even though
+  no task state changed. Its status says that user guidance was requested, but
+  it does not create a user-input boundary.
+- A failed `todo_write` advances the reminder clock. The visible checklist can
+  remain unchanged for ten more turns without a reminder.
+- Small-model budget text says that a todo update resets the phase budget, but
+  the implementation resets only after a context-tree phase transition.
+- Workboard `lastSubstantiveProgress` treats successful discovery reads as
+  progress. This conflicts with the task and completion ledgers, where a read
+  is evidence but not advancement.
+
+## Required invariants
+
+- [x] Every runner binds todo tools to its own immutable session scope.
+- [x] Model-supplied todo session IDs cannot override the host scope.
+- [x] Concurrent runners cannot redirect one another's todo reads or writes.
+- [x] A fresh task archives the prior checklist and clears the active persisted
+  projection so the TUI, runner, REST, and Telegram views agree.
+- [x] Only a confirmed mutation, todo transition, workboard transition,
+  assertive verifier receipt, or delivery receipt counts as authoritative task
+  progress.
+- [x] Reads, searches, runtime-authored blocks, no-ops, and changing errors do
+  not extend the run or re-arm brute-force execution. Entry into the first
+  re-engagement cycle is deliberately not gated on prior authoritative
+  progress: that cycle is the rescue attempt for a run that has only read.
+  If it also advances nothing, re-engagement stops at cycle 2.
+- [x] Repeated stasis creates one latched, model-visible convergence review.
+- [x] The convergence review remains advisory. It does not infer completion,
+  manufacture a blocker, or force a user question.
+- [x] The loop intervention maximum does not silently reset without
+  authoritative progress.
+- [x] Todo reminders advance only after a successful state-changing write.
+- [x] Budget exhaustion text describes the actual reset condition.
+- [x] Workboard progress metadata uses the same authoritative distinction.
+- [x] Todo stagnation compares completed leaves, matching the TUI counter.
+
+## Status
+
+Implemented and verified on 2026-09-03.
+
+| Invariant | Implementation | Test |
+| --- | --- | --- |
+| Immutable runner-owned todo scope | `packages/execution/src/tools/todo-write.ts` `bindExecutionScope` | `todo-store.test.ts` "binds todo tools to a host session that model arguments cannot replace" |
+| Concurrent runner isolation | `AgenticRunner` constructor resolves one immutable `_sessionId`; `registerTool` binds it | `todo-store.test.ts` "keeps concurrent runner-scoped todo writes isolated" |
+| Fresh task clears the active projection | `_hidePriorSessionTodosForFreshTask` writes an empty active checklist after archival | `fresh-task-todo-boundary.test.ts` |
+| Typed authoritative progress only | `packages/orchestrator/src/convergence-progress.ts` | `task-convergence-progress.test.ts` (8 cases) |
+| Latched advisory convergence review | `agenticRunner.ts` convergence review block | `task-convergence-progress.test.ts` "latches one advisory review and clears it only after typed progress" |
+| Loop intervention maximum stays latched | `Math.min(maxInterventions, loopInterventionCount + 1)` | `agenticRunner-context-behavior.test.ts` "latches the loop-intervention maximum instead of cycling it back to one" |
+| Workboard progress uses the same distinction | `recordWorkboardToolCall` returns a status-fingerprint transition | `workboard-run-continuity.test.ts` "does not record a successful discovery read as task advancement" |
+
+Full suites run green: `@omnius/execution` 1507 passed, `@omnius/orchestrator`
+2303 passed (200 files, 2 skipped), `omnius` CLI 2428 passed (258 files).
+Typecheck passes for all three packages and `pnpm -r build` completes.
+
+The reported counter was never attributed to a specific field run. The evidence
+request below stays open.
+
+One correction was made during verification. An earlier draft of this work also
+refused to enter brute-force cycle 1 whenever the primary run recorded no
+authoritative advancement. That inverted the purpose of re-engagement and broke
+eight existing brute-force tests, because a run that has only read is exactly
+the run that still needs one push to act. The cycle-2 gate already bounds the
+thrash this work order set out to stop, so the cycle-1 refusal was removed.
+
+## Deterministic verification
+
+- [x] Scoped todo tools ignore conflicting model-supplied session IDs.
+- [x] Concurrent scoped todo tools write isolated session files.
+- [x] A same-session fresh task clears the old visible todo projection after
+  archiving it.
+- [x] Forty distinct reads produce zero authoritative progress.
+- [x] Changing failures produce zero authoritative progress.
+- [x] Mutations, todo transitions, workboard transitions, assertive verifiers,
+  and delivery receipts produce typed progress.
+- [x] A convergence review latches after the configured unchanged frontier and
+  clears only after typed progress.
+- [x] The loop intervention maximum stays latched instead of cycling to one.
+- [x] Focused execution, orchestrator, and CLI tests pass.
+- [x] Affected package typechecks and builds pass.
+
+## Evidence request for exact field attribution
+
+If the original run is available, collect only a redacted structural slice:
+the exact counter line or screenshot, Omnius version, model/backend, launch
+surface, incident time and timezone, session/run IDs, five turns before the
+stall through ten turns after it, todo IDs/statuses/revisions, workboard card
+statuses, and completion-ledger statuses. Do not copy a complete `.omnius`
+directory because debug previews can contain user text and source fragments.

package/docs/work-orders/telegram-dmn/WO-22-dmn-outreach-and-learning.md

@@ -0,0 +1,85 @@
+# WO-22: DMN outreach, DM sharing, and outcome learning
+
+## Incident
+
+On 2026-09-03 at 17:07 PDT the bot posted in the OMNIUS group without being
+addressed. The operator asked whether this was self-induced reflection.
+
+It was not. The post was reactive to two inbound messages (5680, 5681) and
+authorized as `direct_turn`. The stated router reason — *"Sender reacts to
+bot's recent output in active thread"* — was false: the cited self-bearing
+evidence was `turn:5678`, a bot reply from **09-01**, to a different user,
+about a different topic.
+
+Two independent defects produced that:
+
+- `buildTelegramInteractionEvidencePacket` selects causal history with
+  `.slice(-8)` and **no time bound**. In a quiet room the last 8 messages
+  reach back days.
+- The `direct_turn` validator requires a cited self-bearing evidence entry but
+  checks only *kind* and *actor*, never *recency*. A two-day-old turn satisfied
+  it.
+
+Investigating that surfaced the larger finding below.
+
+## The real finding: reflection could not speak
+
+Omnius reflects constantly — **465 daydream artifacts on 2026-09-03 alone** —
+and roughly two thirds of recent artifacts proposed a same-group follow-up.
+None were ever sent. Across 639 committed routing decisions in this workspace,
+`social_intervention` fired exactly **once**; the other 432 replies were all
+`direct_turn`.
+
+Root cause was ID plumbing, not judgement:
+
+- The extraction schema advertised `reply_to_message_id: 1` and
+  `source_message_ids: [1]` as examples. **78.6% of live follow-up proposals
+  echoed a placeholder-like id**, `1` being the single most common value.
+- Those ids never matched real Telegram ids, so `candidateMessageIds` was
+  almost always empty.
+- The discretion prompt then instructed the model to *"cite at least one
+  current eligible Telegram message ID"* **without ever listing which ids were
+  eligible**, while the host discarded any follow-up citing an id outside that
+  undisclosed set.
+
+Silence was the only reachable outcome.
+
+## Changes
+
+| Area | Change |
+| --- | --- |
+| Extraction prompt | Schema example ids replaced with non-numeric sentinels; explicit rule that ids must appear literally in the corpus; the valid id set is now stated. |
+| Extraction result | `runTelegramReflectionExtraction` drops any `source_message_ids` / `reply_to_message_id` absent from the corpus, so a copied or hallucinated anchor cannot reach the follow-up gate. |
+| Discretion prompt | Now supplies `Eligible evidence message IDs`, the exact set the host validates against. |
+| DM reflection | Idle reflection and outreach admitted for private chats **only after that chat opts in** via `/reflect auto on`. Default remains off. |
+| Outcome learning | New `telegram-outreach-outcomes.ts` records every autonomous send and settles it as engaged / ignored from subsequent room traffic. Settled history is fed back into the next discretion call. |
+
+## Design notes
+
+- **DMs are opt-in, not default.** A DM is one person's inbox. The capability
+  exists; the host does not enable it on the recipient's behalf.
+- **Silence is recorded as a real signal.** An unanswered outreach settles to
+  `engaged: false` after 45 minutes rather than staying unknown, so the
+  calibration block cannot mistake absence of data for success.
+- **The bot's own traffic never counts as engagement.** Only a human message
+  inside the window closes an outcome positively.
+- **Calibration is neutral before it is negative.** With no settled history the
+  prompt says so and asks the model to judge on merit; it raises the bar only
+  once every settled attempt was ignored.
+
+## Not addressed
+
+The evidence-packet recency defect that produced the original false
+`direct_turn` remains open. Bounding causal history by time would push honest
+unaddressed interjections onto the `social_intervention` path, where they are
+visible and tunable — but `social_intervention` currently has **no validation
+rules at all** (compare five checks for `direct_turn`). Both should be done
+together so interjection becomes a governed capability rather than an
+unvalidated fallback.
+
+## Verification
+
+CLI suite: 2442 passed, 0 failed (261 files). Typecheck clean; `pnpm -r build`
+clean. 14 new tests covering id anchoring, corpus allow-list enforcement,
+outcome settlement, calibration wording, torn-ledger recovery, and bridge
+wiring.

package/npm-shrinkwrap.json

@@ -1,12 +1,12 @@
 {
   "name": "omnius",
-  "version": "1.0.689",
+  "version": "1.0.691",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "omnius",
-      "version": "1.0.689",
+      "version": "1.0.691",
       "bundleDependencies": [
         "image-to-ascii"
       ],

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "omnius",
-  "version": "1.0.689",
+  "version": "1.0.691",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/library.js",