EPS
← All tasks·#1828Completed

workflow-fix: harden _breadcrumb_fields against prose-glued/empty key=value tokens

kind: infra#wf-fix#wf-fix-fp:2331b34907b3
track:

Highlight text on any card (body, plan, or an event) to anchor a comment, or leave a whole-task comment here. Mention @claude to summon a reply.

No comments yet.

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a workflow-fix candidate raised on task #1689 (emitting agent: issue-orchestrator).

Goal

Harden _breadcrumb_fields token parsing so prose-adjacent tokens (glued trailing punctuation, bare key= substrings in prose) cannot misroute executing_followup_label to the queue-head fallback.

Workflow gap

  • Bug observed: stage-dispatch breadcrumb label token with glued trailing punctuation (label=wider-lambda-ceilings:) or a prose label= substring re-binding the field to empty makes executing_followup_label fall back to the queue-head unrun label, misidentifying the executing round.
  • Why it is a workflow gap: _breadcrumb_fields whitespace-splits key=value tokens with last-wins semantics and no value hygiene, so (a) a breadcrumb author who ends the token run with a prose colon (label=wider-lambda-ceilings: detached ..., task #1689 event v110, 2026-07-29T08:28:32Z) produces a label that matches no group, and (b) any prose containing a bare key= sequence ("its label= token was prose-glued", #1689 event v115) re-binds the field to EMPTY, silently discarding the correct earlier token. Both make executing_followup_label fall back to the DISPATCHABLE queue head — on #1689 that was the user-chat label derived-vs-free-answer-map, NOT the executing round wider-lambda-ceilings. Downstream consumers of the helper (autonomous_session_watch._post_followup_run_marker's on-behalf run-marker post; the Step 9b step-4 completion-marker label derivation; tick_triage.py's pid-probe crumb read) would then close / probe the WRONG label — a never-run user-chat round would be marked run.
  • Confidence (emitter): high (live repro on #1689 events v110/v115/v116, 2026-07-29: v110 parsed label='wider-lambda-ceilings:' → no group match; v115 parsed label='' via prose re-bind; the equals-free v116 re-post resolved correctly).
  • verified-at-filing: grep -n "def _breadcrumb_fields" src/explore_persona_space/task_workflow.py → 1 hit (line 1907) in the named target; consumers grep grep -rln "_breadcrumb_fields" src/ scripts/ → src/explore_persona_space/task_workflow.py + scripts/tick_triage.py (consumer via the shared helper, no independent parser). Live-repro parse values reproduced at filing time via task_workflow._breadcrumb_fields on #1689 events v110 ('wider-lambda-ceilings:') and v115 ('') (2026-07-29).

Proposed change (candidate diff sketch — refine in planning)

  # src/explore_persona_space/task_workflow.py, _breadcrumb_fields (line ~1907)
+ _BREADCRUMB_VALUE_TRAILING_PUNCT = ":;,."
  ...
-         fields[key] = value
+         value = value.rstrip(_BREADCRUMB_VALUE_TRAILING_PUNCT)
+         if value and key not in fields:
+             # first NON-EMPTY wins: a later bare `key=` prose substring can
+             # no longer re-bind an already-parsed field to empty/garbage
+             fields[key] = value

Planner note: weigh first-non-empty-wins vs last-wins semantics against existing pinned tests (tests/test_task_workflow*.py); trailing-punct strip must not eat legitimate values (labels are kebab-slugs, paths in log= / harvest= never end in :;,. by convention — state this in the plan). Add regression tests reproducing both #1689 shapes (glued colon; prose label= re-bind).

Scope / surfaces

  • Primary target: src/explore_persona_space/task_workflow.py
  • Grep the workflow surface for the pattern before editing (grep -rn "_breadcrumb_fields" src/ scripts/ .claude/) and update every hit; list them in the plan. (scripts/tick_triage.py consumes the shared helper — behavior change flows through automatically; verify its pid-probe arm against the hardened parser.)

Constraints / invariants

  • Workflow-surface only — never experiment code, configs/, or tasks/.
  • scripts/workflow_lint.py --check-asks passes; ruff on touched files passes; if workflow.yaml or CLAUDE.md change, they stay consistent with the rule file.
  • This session runs under EPM_WORKFLOW_FIX_SESSION=1 and carries a workflow_fix_target: Provenance line — it MUST NOT auto-route any of its own subagents' workflow-fix candidates (recursion guard, .claude/rules/workflow-fix-on-bug.md § Recursion guard).

Provenance

  • workflow_fix_target: src/explore_persona_space/task_workflow.py
  • fingerprint: 2331b34907b3

target_file: src/explore_persona_space/task_workflow.py bug_observed: stage-dispatch breadcrumb label token with glued trailing punctuation (label=wider-lambda-ceilings:) or a prose 'label=' substring re-binding the field to empty makes executing_followup_label fall back to the queue-head unrun label, misidentifying the executing round why_workflow_gap: _breadcrumb_fields whitespace-splits key=value tokens last-wins with no value hygiene, so prose-glued punctuation or bare key= prose substrings silently misroute the round-label resolution every consumer (watcher on-behalf run-marker, step-4 completion marker, tick_triage pid probe) depends on proposed_change: harden _breadcrumb_fields token parsing: strip trailing punctuation from values and ignore empty-value rebinds so prose-adjacent tokens cannot misroute executing_followup_label diff_sketch: |

  • value = value.rstrip(":;,.")
  • if value and key not in fields:
  • fields[key] = value
  • fields[key] = value confidence: high related_task: #1689
Activity