Skip to content

Historical field note

This record preserves ground-truth investigation evidence for a single dream finding. It is not a current operating contract; use the numbered book and current runbooks for instructions.

Ground Truth: dream finding dreamrepair:cc1dedb0d3036d289aafc1e42b4a22aa (slack failure_pattern)

Task: Characterize the sole evidence backing a low-confidence dream-repair finding of kind failure_pattern, scope project, project mac, provider label slack, fingerprint dreamrepair:cc1dedb0d3036d289aafc1e42b4a22aa, confidence low (score 0.35). Determine the concrete failure mode, the affected code/config surface (if any), and whether the finding is a recurring pattern or an isolated artifact.

Sole evidence record: mem_843510f5c2344eab881036d529b392a5 (record_type deployment_learning:mac, schema mac.deployment_learning.v1), authored by mac-task-executor.

Originating task: task_2a25617199eb43dd8cf95de6c40ef0a1 — "Convert fleet config to OpenClaw terminology and constrain Slack/persona runtime to openclaw|none" (evidence_type=repo_change).

Prepared by: fleet worker (investigation node; no production code, test, skill, config, or deploy edits).

Verdict: NO LIVE DEFECT — stale preliminary self-evidence; the "slack" label is incidental

The finding rests on a single mid-run, executor-authored self-snapshot that was superseded within the same task. The originating task did not fail: it was reviewed, approved, and published to main, and a second deployment-learning record for the same task records the authoritative outcome approved_published. The slack provider label is derived from the task title/description wording ("Slack/persona runtime"), not from any Slack runtime, chat-gateway, or persona execution failure. This is an isolated, non-recurring evidence-provenance artifact, not an actionable defect.

No files under src/mac/, tests/, skills/, .mac/, or deploy/ were modified by this investigation.

What the sole evidence record actually says

mem_843510f5c2344eab881036d529b392a5 (created 2026-07-26T09:18:19Z by mac-task-executor) carries this mac.deployment_learning.v1 payload:

  • outcome: failure
  • evidence_type: repo_change
  • signals: tests=fail, returncode=0, pushed=false, files_changed=4, checks_pass=null
  • error_signature: not a stack trace or failing assertion — it is a truncated prose recap of the change the worker performed ("Converted the fleet chat-gateway config and inspection surfaces to OpenClaw terminology, added a persona/Slack runtime selector allow-list (openclaw|none) mirroring network.provider, and dropped 'herm…").

Two facts make this a preliminary snapshot rather than a terminal failure:

  1. returncode=0 with tests=fail and pushed=false is the shape the executor records mid-run, before the contract suite has been finalized and before the guarded push — not the shape of a task that terminated in failure.
  2. The error_signature names no offending module, no failing test id, and no reproduction. It is the run's own change summary, so it cannot be used to locate a defect.

The originating task succeeded — the failure was superseded

The task ledger for task_2a25617199eb43dd8cf95de6c40ef0a1 shows a clean success path, all within a single lease:

  • running → needs_review (worker submitted worker-result.json, 8 artifacts) at 09:18:57Z.
  • needs_review → reviewing; an independent reviewer was assigned.
  • Review approved via signed verdict evidence at 09:23:25Z.
  • Canonical branch integration verified; published to git://main (pub_a802c30011fd4705a618afa428e2277f) and transitioned to completed at 13:07:10Z.

Crucially, a second deployment_learning:mac record for the same task, mem_627afb5e4e2c4f09b495eea8984a7b10 (created 2026-07-26T13:07:10Z by hub-review-workflow), records outcome=approved_published (signals.stage=hub_review). The dream classifier consumed only the earlier failure snapshot and did not weigh the later authoritative approved_published record.

The intended change landed and is intact on main

The OpenClaw conversion and the constrained runtime selector this task was about are present in the tree:

  • src/mac/fleet_setup.py reads the openclaw: block with a legacy hermes: read-only fallback and validates the gateway impl against the allow-list — "gateway_impl must be openclaw or none" (src/mac/fleet_setup.py:208), the allow-list check at src/mac/fleet_setup.py:208, and OpenClaw defaults at src/mac/fleet_setup.py:194.
  • src/mac/hermes_config_surface.py, src/mac/worker.py, and .mac/project.yaml are all present.

There is no failing suite, assertion, trace, or reproducer tied to this finding to repair, and no remaining code/config remediation for the mac chat-gateway / persona / Slack runtime surface.

Is the "slack" provider label causal or incidental?

Incidental. The provider/area label attached to the finding comes from the task title and description text — "constrain Slack/persona runtime" — which the nap/dream summarizer echoes verbatim into the finding's observations and summary. Nothing in the evidence points to a Slack transport, chat-gateway runtime, or persona-binding failure. The word "Slack" appears only because the task was about renaming and constraining that config surface, not because that surface failed.

Recurring pattern or isolated event?

Isolated. The dream failure_pattern record is explicitly "Supported by 1 memory record(s)" — a single deployment_learning:mac evidence item (record_type_counts: {deployment_learning:mac: 1}). The 0.35 confidence is the classifier's deterministic single-record structural floor (an evidence-volume signal), not a corroborated defect. Multiple later dream:failure_pattern records exist for this task, but they are re-emissions of the same nap window over the same single evidence record, not independent occurrences.

Evidence-provenance gap (the real, generic issue)

The finding is a false positive produced by a provenance gap, not a product defect: the dream/nap failure-pattern classifier keyed on an executor preliminary deployment_learning self-snapshot (outcome=failure, tests=fail, pushed=false) and did not reconcile it against the terminal record for the same task_id (outcome=approved_published). A durable fix belongs to the classifier/consolidator (prefer the latest terminal outcome per task_id, or suppress preliminary failure snapshots once a later approved_published record exists), and is intentionally out of scope for this investigation, which was directed not to change any skills, tools, or config.

Reproduction (read-only, hub queries)

  • mac task show task_2a25617199eb43dd8cf95de6c40ef0a1 --json — shows the approved/published/completed path and the git://main publication.
  • mac admin memory search --task-id task_2a25617199eb43dd8cf95de6c40ef0a1 --json — shows both deployment_learning:mac records (preliminary failure and terminal approved_published) plus the single-evidence dream:failure_pattern records that cite only the former.