Operational Synthesis — 2026-09-04 10:11Z

• status: complete
• scope: scheduled higher-level operational synthesis
• material_digest: 4f47a542f6289c39c293e906b57118494997435350891aeacdb68b78b881fa4a

Summary

Material change is present: the user resumed activity around the NCN fallback/travel weekend and has an active architecture request for a durable, session-independent polling control plane. Runtime and memory workers are currently healthy, but several long-lived operational gaps remain unresolved.

Evidence reviewed

• model-conversation-log.jsonl: 6,386 rows; latest user turn 2026-09-04T08:15:25Z and assistant turn 08:15:59Z. Recent topic: additional Leipzig cultural events after the user was too tired to go out.
• memory-heartbeat.json: restore successful, 7,520 entries; model-conversation source is current through runtime rows and semantic_changed=true.
• model-heartbeat.json: 2026-09-04T10:01:56Z, rc=0, acpx_rc=0, NO_CHANGE.
• context-guard.json: direct ACP normal at 265,251 provider tokens (25.3% of advertised 1.05M), no compaction and no reload recommendation; live model is openai/gpt-5.6-terra.
• continuity-summary.txt: regenerated 10:08Z but still incorrectly says kimi-for-coding/k2p6.
• semantic-candidates.jsonl: one new candidate is the transient Friday cultural-events request; it should remain review material, not durable memory. Older candidate rows are predominantly stale. memory-candidates.jsonl still has 13 candidate-status rows for curator review.
• actions.jsonl / notices.jsonl: durable open work remains, including worker model repair, decision-state refresh, semantic-memory cleanup, provider-ceiling proof, notice-source hardening, and the newly active polling-control-plane design.
• directed/latest.md is stale (no directed report since June); directed/index.json was refreshed but contains only old directed reports.

Findings

1. Runtime healthy, telemetry incomplete. Memory restore and model heartbeat succeeded, and the live ACP context is comfortably normal. This does not prove the provider ceiling or unattended control-plane health.
2. Current user context is active and practical. NCN appears unavailable/sold out; the user is considering Leipzig/Nuremberg/Dresden/Berlin alternatives. Event recommendations are time-sensitive and must be re-verified live before reuse. No durable-memory promotion is warranted from the transient cultural-events request.
3. Polling architecture is the highest-value open engineering thread. The user explicitly wants arbitrary reasonable schedules, durable watches/queueing, dedupe/debounce, retry/backoff, deadlines, wake budgets, shadow mode, and visible subagent lifecycle. The agreed safe direction is an isolated service before the native backend: poller → durable queue → (transitional ACP-resume adapter) → later prompt-builder/fresh backend. Do not evolve the existing ACP-coupled timer dispatcher in place.
4. Generated model identity remains stale. The live context and heartbeat are GPT-5.6 Terra while the generated continuity summary names unavailable Kimi K2.6. Repair must target the scheduled-worker source/configuration and then prove two clean runs; do not edit the generated summary directly.

Hypotheses

• The recent semantic_changed=true reflects extraction of the new transient events request rather than a durable preference; curator should mark it ignored/stale through the normal candidate helper after review.
• The repeated stale model identity likely originates in scheduled-worker configuration or generator fallback, not current direct ACP state. This remains unverified until source inspection/control-plane smoke tests are permitted.

Actions / backlog

• Main agent/infrastructure: repair scheduled-worker model/config selection and verify two consecutive clean runs.
• Main agent/infrastructure: stage the isolated polling control plane in shadow mode, beginning with subagent completion, service health, and URL/inbox watches; preserve the narrow ACP adapter only as a transition.
• Main agent/infrastructure: expose timer, subagent, approval, observability, and backlog status through a reliable control-plane surface; independently prove provider ceiling before changing guard thresholds.
• Curator/main agent: review the 13 legacy memory candidates and audit low-signal fragments in semantic-memory.json; do not treat candidates as facts.
• Main agent: refresh stale decision-state.json, and separately harden directed/index notice regeneration so resolved/obsolete reports are not re-injected.

Curated-memory proposals

None. The new events discussion is transient and location/date-specific. The polling-control-plane requirements are already captured as an open action and recent topic, not a personal memory.

Warm-memory update

Updated recent-topics.md with the 2026-09-04 runtime state, current Leipzig/NCN context, polling architecture priority, stale worker-identity warning, and verification gap.

Missing tools / blockers

• Matrix-lite native timer/subagent/status MCP tools were not exposed in the available tool manifest this cycle; no independent timer, subagent, approvals, observability, or persistent_chat_log_backlog_status check was possible.
• turns.jsonl is referenced by the model-conversation source but was not directly inspected here; no backlog inference is made from canonical-vs-model row counts.
• No native or async subagent was used; the earlier architecture-review job 9cec5ce3080c remains a known timeout/non-result and is not evidence against the design.

Operational risks / robustness suggestions

• Prioritize source-level worker model repair; otherwise scheduled synthesis can silently regress to ProviderModelNotFoundError despite a healthy main session.
• Add a generated-state provenance/check: continuity summary should record the actual resolved model and source, and fail visibly on mismatch rather than publishing stale fallback identity.
• Keep notice regeneration idempotent by deterministic source IDs and filtering resolved/obsolete reports. Avoid routine pruning in this cycle because notices are append/history artifacts and no safe helper was exposed.
• Preserve the no-reload policy: current context is normal, so do not replace the live persistent backend.

Files updated

• This report (checkpoint replaced with completed synthesis).
• /home/matrix-lite/state/persistent-chat/recent-topics.md (warm-memory current-topic update).

No JSONL findings/hypotheses/actions/notices were appended; existing open records already cover the durable issues without duplication.