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:4becfa8d (slack failure_pattern)

Finding: dreamrepair:4becfa8dfd2a06539d4dbb0fb4f53be9, kind failure_pattern, scope project, project mac, confidence low, evidence_count=1, provider label slack.

Candidate summary (as supplied): "failure pattern for task=task_4b603d746c2a4359b2c6b5c2ef2ac653 project=mac. Supported by 1 memory record(s): [failure] Audit slack dream finding evidence and provenance (repo_change)".

Sole cited evidence: mem_ad60e715d76a41199f256975851d6df8 (record_type=deployment_learning:mac).

Prepared by: fleet worker (read-only investigation). No files under src/, tests/, skills/, or deploy/ were modified.

Verdict: NOT ACTIONABLE

The attached evidence does not name a reproducible Slack defect. The slack token is a word in a prior audit task title that the nap dreamer copied into an observation. The sole supporting record is a deployment_learning:mac closure-shaped snapshot for that audit task, with a blank error signature. Hub lookups that would retrieve the memory row and originating task were unavailable in this sandbox (policy_denied / hub-mode gaps). Nothing in the current tree implicates a Slack transport, skill, or API as the failing surface.

Q1 — Actionability

Does the attached evidence name a reproducible Slack defect (named Slack tool/skill/API/binding, error signature, failing test, dated incident)?

No.

The observation string is exactly the shape _record_observation emits for mac.deployment_learning.v1 when error_signature is empty (src/mac/nap_consolidator.py):

  • [outcome] title (evidence_type)
  • the suffix failed with <error_signature> is appended only when error_signature is non-empty after strip.

The supplied observation is [failure] Audit slack dream finding evidence and provenance (repo_change) with no failed with … suffix. That is a blank error signature, not a stack, HTTP status, Slack API method, or test id.

Named Slack surfaces in this tree

Path named in the task Present? Implicated by this finding?
src/mac/_hermes/gateway/platforms/slack.py absent (no _hermes/ tree) no
src/mac/_hermes/hermes_cli/slack_cli.py absent no
src/mac/notifier_service.py present; slack is a supported channel type no — not named in the evidence
src/mac/communication_service.py present; slack is a channel name in CHANNELS no — not named in the evidence
scripts/mac-fetch-slack-secrets.py present (vault → local Hermes env upsert) no
scripts/slack-vault-loader.py present (local tokens → vault) no
tests/test_slack_*.py tests/test_slack_secrets_fetcher.py only no failing test is cited

Related but not in the named list: tests/test_hermes_config_surface_slack_tokens.py covers token promotion into the Hermes config env block. It is likewise not referenced by the finding.

Named evidence gap: there is no Slack module, skill, API method, binding, error signature, failing test identifier, or dated incident in the attached evidence. The only Slack signal is the English word slack inside a prior task title that the default dreamer copies verbatim.

Q2 — Provenance

Is the single supporting record self-referential (closure/outcome memory of a prior task in this same investigation lineage) rather than an independent Slack failure signal?

Yes, as far as the producer code and the attached strings can prove. Hub confirmation of the memory row itself was unavailable.

Hub lookups (attempted, not inferred)

Read-only mac CLI against the configured hub, --json where applicable:

  • mac task show task_4b603d746c2a4359b2c6b5c2ef2ac653 → hub GET /tasks/<id> not permitted by policy.
  • mac admin memory search --task-id task_4b603d… and --record-type deployment_learning:mac → GET /memory?… not permitted by policy.
  • mac admin dream list → list_dream_runs not yet supported in hub mode.
  • mac admin memory recall-dreams --kind failure_pattern … → GET /v1/memory/dreams/recall?… not permitted by policy.
  • Direct GET /health on the hub URL → not permitted by policy.

Those lookups are therefore unavailable, not used as positive evidence.

What the producer code does prove without the hub

NapConsolidatorService._group_records keys groups on the record's own task_id (and the task's project). _group_label renders task=<task_id> project=<project>. _default_dreamer titles the artifact "<kind words> for <group_label>" and prefixes the summary with that title plus Supported by N memory record(s): and up to three _record_observation lines.

The supplied candidate summary matches that template for task_4b603d746c2a4359b2c6b5c2ef2ac653 / project mac / one observation. A candidate labelled that way can only be built from records whose task_id is that id.

record_type=deployment_learning:mac is the closure lesson type written by:

  • build_learning_record / record_deployment_learning in src/mac/executor_memory.py (created_by=mac-task-executor);
  • the hub project-failure lesson in src/mac/services.py (created_by=mac-hub-review, outcome=failure).

The observation title is an audit of Slack dream-finding evidence and provenance with evidence_type=repo_change — the same investigation lineage, not a Slack send/auth/webhook incident. Combined with a blank error signature, the sole evidence is the prior audit's own failure-shaped learning record, not an independent Slack runtime signal.

Q3 — Live producer

Which stages of the documented regeneration loop still exist?

Confirmed starting points were re-checked in this tree:

  • src/mac/nap_consolidator.py is present. It still renders prior-task titles into dream candidates via _record_observation, _dream_kind, _confidence_for_records, and _default_dreamer.
  • src/mac/dream_scanner.py, src/mac/dream_cycle_classifier.py, and src/mac/dream_repair_tasks.py do not exist. src/mac/dreaming/__init__.py states they were replaced by mac.dreaming (asynchronous candidate-store curation, not a keyword scanner that files repair tasks).
  • A search for the string dreamrepair under src/, scripts/, tests/, deploy/, and skills/ returns zero matches.

Candidate manufacture — live (manual / API path)

NapConsolidatorService.consolidate_agent still defaults emit_dream_artifacts=True. When that flag is true it writes dream:<kind> memory rows from _default_dreamer (or a caller-supplied dreamer_fn).

That path is still reachable:

  • HTTP POST /agents/{agent_id}/nap-consolidate (src/mac/api.py), body default emit_dream_artifacts=True.
  • CLI mac admin nap consolidate (src/mac/cli.py), dreams on unless --no-dreams.

Single-record groups still get confidence low / score 0.35 (_confidence_for_records). _dream_kind still returns failure_pattern if the concatenated record type/content contains failure, failed, or error — including a deployment-learning outcome of failure.

Candidate manufacture — scheduled path mitigated

ControlPlane.run_nap_cycle always calls consolidate_nap with emit_dream_artifacts=False, with an inline comment that the legacy dreamer is replaced by mac.dreaming. The systemd unit deploy/systemd/mac-nap-tick.service runs mac admin nap cycle, so the ticker does not emit the old dream:failure_pattern rows.

If run_nap_cycle is invoked with emit_dream_artifacts=True, it runs run_dream_cycle (the new mac.dreaming pipeline), not _default_dreamer.

Provider labelling / dreamrepair: fingerprinting — gone

No live module assigns area_name=slack or mints dreamrepair: keys. Those behaviours lived in the removed classifier / repair-task modules. mac.dreaming does not reproduce that labelling.

Repair-task filing — gone

run_nap_cycle still returns keys repair_tasks and repair_task_error. They are initialized empty/None and never written. The comment in src/mac/services.py states the low-confidence repair-task filer is gone.

Reopen criteria

Reopen this finding as a Slack repair only if all of the following become true:

  1. Independent evidence names a concrete Slack surface that still exists (module, CLI, notifier/communication binding, secret-fetcher path, or a failing test id), not merely the word slack in a task title.
  2. That evidence includes a non-blank error signature, reproduction, or dated incident — not only a deployment_learning:* closure of an audit / dream-repair investigation.
  3. Hub (or equivalent ledger) lookup of the cited memory id succeeds and shows a payload that is not the prior lineage task's own outcome snapshot.

Until then, treat dreamrepair:4becfa8dfd2a06539d4dbb0fb4f53be9 as not actionable. Residual risk is the still-reachable manual/API nap-consolidate path (emit_dream_artifacts=True by default), which can re-emit title-echo dream:failure_pattern rows, but it cannot mint dreamrepair: fingerprints or file repair tasks in this tree.