EPS Dashboard

Overview / Motivation

Auto-filed by the /daily 2026-07-26 problem sweep (route 2). Surfaced by 5 independent miner group(s) over the 2026-07-26 session transcripts.

Goal

Extend the planner's grounding discipline from hyperparameters to the four grep-answerable claim classes that failed on 2026-07-26 — tool behavior, symbol existence, cited line numbers, and task-body-named required edits — and require tracing the exit path of any shell control flow a plan embeds.

Workflow gap

Bug observed: Seven grounding failures across five sessions on 2026-07-26, each costing a plan-revision or fact-checker round: a wrong line number for a symbol's second call site, a nonexistent library API plus a wrong file citation, two independent false claims about workflow_lint.py's no-flags bundling, a silently dropped required doc edit, an acceptance criterion its own instrument could not measure, and a bare false in embedded shell that does not halt the enclosing block.
Why it is a workflow gap: .claude/agents/planner.md §11 requires a Source: line for load-bearing HYPERPARAMETERS only; nothing extends that bar to a claim about what a repo script does, whether a symbol exists, or which line a call site sits on — and no §4 pre-return check asserts that every edit the task body names verbatim appears in the plan.
Confidence (emitter): high
verified-at-filing: absence probes against BOTH the spec and the section reference it defers to — grep -ci -- '<pat>' .claude/agents/planner.md .claude/rules/planner-section-reference.md for symbol-existence, grep -rn, tool-behavior, line number, scope shrinkage, control flow, exit path → 0 hits in both files for all seven patterns; acceptance criterion → 2 hits in planner.md (L602-604), both inside the count-style self-count rule, 0 in the reference. §11 (L511-523) scopes Source: to "one entry per load-bearing hyperparameter"; §12 (L525-557) requires Confidence/Source/How-to-verify but no verbatim grep output; §4 (L333-350) lists hard-requirement items with no task-body-edit enumeration; §7 (L396-408) requires gates be "jointly satisfiable" and precedent-coherent but never names the measuring instrument (2026-07-26)

Evidence

Session 7df6ce4c, 09:48:27Z: plan §12 assumption A1 asserted recommended_timeout_s has call sites at L1669 and L466; the fact-checker found the second at L1725 — "**A1: WRONG.** recommended_timeout_s has TWO active call sites, but they are at **L1669** (--map-files branch — CORRECT) and **L1725** (the diff-path/invariant-set …". Cost: one fact-check finding plus two plan-patch turns, the second of which itself failed.
Session 06447a89, 07:21:55Z: the fact-checker returned HAS_WRONG on 2 of 11 assumptions — "§4 pseudocode uses task_workflow.load_registry — actual API is private _load_registry … §12 assumption 12 cites SKILL.md files for the phrase — actually only in CLAUDE.md line 119." The plan had itself flagged the API at MEDIUM confidence and deferred it to "the implementer should inspect the exact API surface at implementation time" rather than grepping it. Cost: 4 plan_patch.py invocations.
Session 06447a89, 07:29:47Z: task #1691's body (lines 87-88) required TWO CLAUDE.md edits; plan v5 §4.8 named only one. Three lens critics APPROVEd; only the consistency-checker caught it — "the plan §4.8 names ONLY the § Autonomous-session watcher edit and OMITS the § Codex ensemble review edit. This is a scope-shrinkage MISMATCH." Cost: one extra plan revision (v5→v6); a half-fix would have shipped without that one reviewer in the batch.
Sessions c0319d9e (11:45:47Z) and 7ce3a81f (07:23:48Z), independently: both plans asserted that workflow_lint.py's no-flags default run bundles --check-asks; it does not (main() gates it on if args.check_asks: alone). "the plan Assumption 5 said --check-asks IS bundled — that specific detail is **WRONG**". The same claim class, wrong the same way, in two unrelated plans on one day — asserted rather than grepped.
Session c0319d9e, 11:55:39Z: the Statistics & Measurement critic returned REVISE because §7 criterion 2 claimed byte-identical anchor-sentence enforcement while §4.2 specified a COUNT check on 30-char anchor prefixes — "criterion 2's automation claim is FALSE and criterion 2 is not mechanically enforced. This is exactly the acceptance-vs-instrument mismatch the Statistics & Measurement lens exists to catch". Cost: one Phase-3 revision round.
Session 35d7c0fa, 09:25:23Z: plan v1 for #1694 gated a push behind a diff-verify fence whose failure arm ended in a bare false — "Edit A's terminal falsein the diff-verify fence does NOT prevent the immediately-following push block … from running and overwriting the sentinel withlanded." The plan as written would have shipped the exact silent-success bug the task existed to fix; caught at methodology-critic round 1, ~7 min for the extra round.

Proposed change

§11: extend the Source: bar beyond hyperparameters to TOOL-BEHAVIOR claims — any assertion about what a repo script, lint, or CLI does under a given invocation carries a Source: naming the grep or file:line actually read, and the planner RUNS the one-line mechanical check (--help, a grep of the flag's main() chain) rather than reading the pattern by eye.
§12: any assumption naming a line number quotes the grep -n '<symbol>' output verbatim in the row. A line number written without its grep output is not grounded.
§4 (or §12, whichever fits the byte budget): forbid deferring a grep-answerable symbol-existence question to the implementer — every module.symbol written into plan pseudocode is confirmed at plan time with a recorded grep -rn 'def <symbol>' src/ scripts/; a symbol that does not resolve is rewritten, not shipped at MEDIUM confidence.
§4 pre-return self-check: before returning, enumerate every file + section the task body names as a required edit and assert each appears in §4 Design; list any deliberate omission with a one-line reason. This is the consistency-checker's catch made a planner-side duty.
§7: state, per acceptance criterion, WHICH §4 mechanism measures it and what that mechanism actually compares (count vs equality vs presence). The existing L602 self-count rule covers count-style criteria only.
Plan-quality: when a plan embeds shell control flow, trace the exit path of every failure arm and state what halts the enclosing block — a bare false or exit inside a branch does not stop a sibling block; prefer an explicit OK=yes|no variable gate over &&/exit-status chaining.
BYTE BUDGET — binding placement constraint: .claude/agents/planner.md is 38,368 B against AGENT_SPEC_FAIL_BYTES = 40,000 (scripts/workflow_lint.py L11740) and is deliberately NOT grandfathered (the AGENT_SPEC_SIZE_GRANDFATHER header names planner.md and critic.md as excluded, #838). Headroom is 1,632 B for all six additions. Put the full templates in .claude/rules/planner-section-reference.md (75,176 B, the on-demand file §4/§7/§11/§12 already defer to) and keep the spec-side text to one compressed clause per section.

Scope / surfaces

Primary target: .claude/agents/planner.md
.claude/rules/planner-section-reference.md (§4 / §7 / §11 / §12 full templates — the byte-budget-safe home for the expanded text)

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

sha-verify (filing-time, #1467): 35d7c0fa cited in commit context does NOT resolve as a commit in this repo at filing time — treat as a transcript/session reference, not a commit.
fingerprint: 29418af1fe8a
workflow_fix_target: .claude/agents/planner.md
fingerprint: PENDING

/daily 2026-07-26 route-2 filing. Miner refs: E-P5, E-P6, C-P9, C-P10, I-P11, H-P10, G-P13.