Skip to content

plat 001

github-actions[bot] edited this page Sep 20, 2026 · 1 revision

← Pulse platform issue index

Internal reconciliation — 2026-09-05

PUL-7B849CC6: keyed-input parser, unknown-ID, sibling/map isolation, explicit-input precedence and opening-prompt tests pass on current source.

Resolved in SQLite for internal tracking with previous concern/detail records preserved in resolution events. Source/tests verified; deployed replay and historical business/module-result repair are not claimed. Full mapping: remaining-report audit.

PLAT-001 — run_full_workflow drops keyed human input

Coordination Value
Assigned agent Claude Code
Ticket state runtime_reverify
Last synchronized 2026-08-04

Claim this ticket in this file before implementation. During active work, update this fragment rather than the shared index; synchronize the index once at handoff, review, or completion.

  • Priority: P0
  • Owner: workflow orchestration/handoff
  • Source finding: HARNESS-RUN-FULL-WORKFLOW-HUMAN-INPUT-LOSS
  • Source database: Workflow/upwork/db/db.sqlite
  • Recorded state: external_action_required, severity high
  • Problem: a human-input override supplied to run_full_workflow did not appear in the target child step's opening prompt.
  • Impact: run-specific safety or scope constraints can be silently ignored while the workflow continues and reports success.
  • Evidence: the saved scheduler conversation contains the two-feed override; the child search-find-and-shortlist session does not; its raw artifact shows that most_recent was scraped anyway.
  • Implementation (2026-08-03): run_full_workflow now parses human_inputs strictly, rejects malformed and unknown step IDs before launch, copies the map into immutable execution/batch contexts, and derives one isolated context per dispatched step. The keyed value reaches regular, message-sequence, todo, and human-input steps without leaking to siblings; an explicit execute_step(human_input=...) remains higher priority.
  • Verification: focused tests cover both MCP-decoded map shapes, fail-closed parsing, unknown IDs, two distinct child values, sibling isolation, map-copy isolation, and explicit single-step precedence. A real scheduled full-run replay remains required before closing the Upwork finding.
  • Current workaround: compare each producing child's prompt and raw provenance with the requested override.
  • Acceptance: an E2E passes distinct keyed values to at least two child steps and proves each child sees only its intended value; missing or unknown keys fail before execution.

Clone this wiki locally