Historical field note

This record preserves prior investigation or planning evidence. It is not a current operating contract; use the numbered book and current runbooks for instructions.

Investigation: dream finding dreamrepair:5404b15fffa355d739c21e138c5cc122 (skill subsystem)

Task: Establish the ground-truth state of the mac skill subsystem so the low-confidence dream-cycle finding can be assessed. No behavior changes. Finding: kind failure_pattern, scope project, fingerprint dreamrepair:5404b15fffa355d739c21e138c5cc122, confidence low, backed by a single evidence record. Sole evidence record: mem_7d29b6dcefb740bfa647da2aa4325023 (deployment_learning:mac) — itself a prior failure to repair a low-confidence skill finding for parent task task_ad368daa48f14b0cbb82795ff833ed5b. Repo areas mapped: src/mac/skill_auto_repair.py, src/mac/dream_repair_tasks.py, src/mac/dream_scanner.py, src/mac/nap_consolidator.py, and skill modules under src/mac/_hermes/ (agent/skill_commands.py, agent/skill_preprocessing.py, tools/skills_guard.py, tools/skills_hub.py, tools/skills_sync.py). Investigated by: fleet worker (investigation only; no production code edits).

Status: NOT ACTIONABLE as a skill-subsystem defect

There is no reproducible skill-subsystem defect behind this finding. The skill-focused test suites all pass on the current tree, and the finding's shape is fully explained by the dream/nap consolidation heuristics classifying a single self-referential "repair failed" memory as a low-confidence failure_pattern. The prior repair attempts recorded against this task line failed at coding-agent sandbox preflight (probe_failed, executor_failed), which is executor/infrastructure behavior, not a skill defect.

Ground Truth Observed

Measured in the task-owned worktree with the bootstrapped .venv (python3 3.12.13; git/gh present; pytest/coverage installed by python3 scripts/bootstrap-project.py). All skill sources were read, not modified.

Skill-focused tests: all green

Run via .venv/bin/pytest:

Test file	Result
tests/test_skill_auto_repair.py	30 passed
tests/test_fleet_skills.py	2 passed
tests/test_hermes_skills_dedup.py	14 passed
Total	46 passed, 0 failed

The vendored Hermes snapshot is present (hermes_vendor.is_vendored() is True), so tests/test_hermes_skills_dedup.py executed for real (the skipif guard did not skip) — its 14 assertions genuinely ran and passed. No skill test currently fails.

Component mapping vs. the fingerprint

The fingerprint is produced by repair_fingerprint() in src/mac/dream_repair_tasks.py, hashing normalized {kind, scope, project, signature, summary, affected} of a dream candidate. It is a dedupe key for a dream candidate, not a pointer into any single skill module. The candidate it keys is kind=failure_pattern, scope=project — the generic bucket the nap consolidator assigns, not a specific defect location.

• src/mac/nap_consolidator.py _dream_kind() returns failure_pattern for any record whose text contains failure / failed / error. The sole evidence record is a prior repair failure, so its text trivially matches and is bucketed as failure_pattern.
• src/mac/nap_consolidator.py _confidence_for_records() returns low for a single supporting record (support < 2). With exactly one evidence record, the finding is necessarily low confidence.
• Together these explain the finding's kind/scope/confidence entirely from the consolidation heuristics acting on one self-referential memory — no signal that a skill module misbehaves.

Skill-subsystem role reference (read-only)

• src/mac/skill_auto_repair.py: guarded staging path for high-confidence dream skill findings (allowlist to skills/ + deploy/skills/, evidence gate, secret/identity scrubber, fleet-generic constraint). It only engages at overall_confidence=high; a low-confidence finding never reaches it. It is currently exercised only by tests/test_skill_auto_repair.py and is not yet wired into a live pipeline call site (no non-test importers found).
• src/mac/_hermes/agent/skill_commands.py, src/mac/_hermes/agent/skill_preprocessing.py: in-agent skill command surface and preprocessing (behavioral, unrelated to dream repair).
• src/mac/_hermes/tools/skills_guard.py, tools/skills_hub.py, tools/skills_sync.py: skill guard/hub/sync tooling. remove_shadowed_top_level_duplicates() in skills_sync.py (the "Ambiguous skill name" dedup fix) is covered and green in tests/test_hermes_skills_dedup.py.

Real defect vs. infrastructure failure

• (a) Real reproducible skill-subsystem defect — none observed. All 46 skill-focused tests pass; the guard/dedup/auto-repair paths behave as specified.
• (b) Executor/preflight infrastructure failure — the prior attempts on this task failed at coding-agent sandbox preflight (probe_failed, executor_failed), i.e. the investigation never started, not that a skill test or skill guard failed. This is an executor/harness condition, outside the skill subsystem, and is not evidence of a skill defect.

Evidence-gap assessment

The finding rests on a single, low-confidence, self-referential record: the sole evidence is a prior failure to repair a low-confidence skill finding, so the "pattern" is one report about a repair not landing rather than an observed skill malfunction. There is no independent corroboration, no failing test, and no reproducer. Per the consolidation heuristics above, this is the expected low-confidence artifact of one "repair failed" memory, not a validated defect.

Determination

• Do not open a skill-behavior repair from this finding: it is not actionable as a skill-subsystem defect.
• The finding should be treated as low-confidence / unsubstantiated and can be aged out or superseded once the executor-preflight failures on the task line are resolved.
• If any follow-up is desired, it belongs to the executor/preflight layer (fixing coding-agent sandbox probe_failed / executor_failed), not to the skill modules.

Reproduction

python3 scripts/bootstrap-project.py
.venv/bin/pytest tests/test_skill_auto_repair.py tests/test_fleet_skills.py tests/test_hermes_skills_dedup.py -q
# => 46 passed