daily-fix: harden the /daily miner brief (JSONL traps, probe

Overview / Motivation

Auto-filed by the /daily 2026-07-26 problem sweep (route 2). Surfaced by 1 independent miner group(s) over the 2026-07-26 session transcripts.

Goal

Give the /daily problem sweep a durable miner-brief composition recipe carrying the three JSONL-reading traps miners keep hitting, and require every miner suggested fix to state whether its named mechanism was probed or inferred.

Workflow gap

• Bug observed: miners spawned by the /daily problem sweep hit three JSONL-reading traps the brief does not warn about — transcript rows with no message key (KeyError: 'message'), multi-file JSONL loops exceeding the default 2-minute Bash timeout, and git calls made after a cd into ~/.claude/projects/... (not a git repo, exit 128) — and miner suggested fix lines name root-cause mechanisms the miner inferred rather than probed, three of which were refuted at filing-compose time.
• Why it is a workflow gap: .claude/skills/daily/SKILL.md has no miner-brief composition section at all, so the brief is re-improvised from scratch each run and every lesson learned in one night's fan-out is lost by the next.
• Confidence (emitter): high
• verified-at-filing: absence greps against the named target — grep -niE 'miner|CHEAPLY|fan.out' .claude/skills/daily/SKILL.md → 1 hit (L770), and it is not a recipe — the word "miners" appears once, in the headless-mode paragraph ("file's content depends on (transcript miners, a driver whose filed #ids …"); grep -niE "get\('message'|KeyError|timeout ≥|not a git repo" .claude/skills/daily/SKILL.md → 0 hits for all three gotchas; grep -niE 'probed|inferred|ran the probe' .claude/skills/daily/SKILL.md → 0 hits. For contrast the equivalent duties DO exist one layer downstream — the filer's verified-at-filing mandate and the unverified-premise labeling convention are both in the same file (L485), which is why the refuted premises were caught at compose time rather than after filing. Landed-fix check: git log --oneline --since='7 days ago' -- .claude/skills/daily/SKILL.md → 6 commits, none adding a miner-brief recipe. (2026-07-26)

Context binding — one target detail corrected. The mined report proposes the change in .claude/skills/daily/SKILL.md "§ Problem sweep, miner-brief composition". The § Problem sweep section exists (L194); the "miner-brief composition" sub-section does not. The change therefore CREATES that sub-section under § Problem sweep. The "How to read the transcripts — CHEAPLY" recipe the miners actually ran against lives only in the per-run ad-hoc brief (logs/daily/mining-2026-07-26/BRIEF.md L25-33) — which is exactly the durability gap: that file is regenerated nightly and inherits nothing.

Evidence

• Session c0a2df1b, subagent miner-interactive-B (agent-a69ace10621b8fd5c), within 06:30–06:50Z: 3 of that miner's 27 tool_result firing events were errors, all from gaps the recipe does not cover — "Exit code 143\nCommand timed out after 2m 0s" on a for f in <6 session ids> loop over multi-MB JSONLs; "Traceback (most recent call last): File \"<string>\", line 7, in <module> KeyError: 'message'" (rows of type queue-operation, attachment and last-prompt carry no message key); "Exit code 128" from a git log -- <paths> run with cwd inside ~/.claude/projects/.... Measured cost: 3 retries, roughly 2.5 min. The other 5 miners in the same fan-out had 0 errors, so this is guidance drift rather than a systemic failure.
• Session c0a2df1b, 2026-07-26T06:36:06Z and 06:44:10Z: three route-2 filings drafted from miner reports were refuted by the filer's own probes before filing — (a) "the --map-files map applies no workflow_lint surcharge", refuted by running recommended_timeout_s (2580 s; the real cause is that the surcharge key is an EXACT path while the map holds 25 sibling files); (b) "code-reviewer PASSed with 7 open BLOCKERs", refuted — code-reviewer.md rule 11 prescribes exactly that shape; (c) "RunPod bootstrap hard-sets BRANCH=main", refuted — threading exists and main is the fallback, so the real defect is an empty repo_branch upstream. Evidence: "C4's premise is mis-diagnosed — the surcharge isn't missing, its key just doesn't match. Verifying C1 and C5, then composing corrected bodies." and "One filing's premise just failed verification — code-reviewer.md rule 11 *designs* PASS-with-persisted-BLOCKER. Dropping that half."
• Measured cost: roughly 10 min of re-verification and body rewrites, and one filing half dropped. Nothing wrong shipped — the filer's discipline held — but each refutation was a round the miner→filing handoff could have avoided by marking the diagnosis as inferred.

Proposed change

• .claude/skills/daily/SKILL.md § Problem sweep — add a miner-brief composition sub-section that is the durable source for the per-run brief, carrying the transcript-reading recipe currently re-improvised each night plus these three lines:
  • rows without a message key exist (queue-operation, attachment, last-prompt) — use .get('message', {}), never row['message'];
  • any multi-file JSONL loop passes an explicit Bash timeout ≥ 300000 ms (the default 2-min cap kills a 6-transcript loop mid-pass, exit 143);
  • run every git call from the repo root — the transcript dirs under ~/.claude/projects/... are not a git repo (exit 128).
• Same sub-section — extend the miner output schema with a probed-vs-inferred field on suggested fix: when the fix names a mechanism (a missing surcharge, a hard-set variable, a rule violation), the miner states probed: <the exact command it ran> or inferred — not probed. A miner is not obliged to probe; it is obliged to say which it did, so the filer knows which premises need a verification round.
• Keep the existing evidence-counting discipline (L214-243) as-is and reference it from the new sub-section rather than restating it.

Scope / surfaces

• Primary target: .claude/skills/daily/SKILL.md
• none

Constraints / invariants

• Workflow-surface only — never experiment code, configs/, or tasks/.
• uv run python scripts/workflow_lint.py passes (no-flags); ruff clean on touched files.
• This session runs under a workflow_fix_target: Provenance line — it MUST NOT auto-route its own subagents' workflow-fix candidates (recursion guard).

Provenance

• fingerprint: 4800548f9ba8

• workflow_fix_target: .claude/skills/daily/SKILL.md

• fingerprint: PENDING

/daily 2026-07-26 route-2 filing. Miner refs: C-P5, C-P6.