Multi Lane Report Assembly

Multi-lane report assembly

Assemble ONE evidence-graded record (or report) from several parallel lane outputs — the standard shape when you fan out research/review lanes and integrate their findings yourself. The lanes produce finding blocks; you lift them verbatim, consolidate, renumber, and gate the result.

When this fires

• You dispatched 2+ lanes that each returned ### F# — <claim> [GRADE] blocks (evidence-first shape) and you must merge them into a single record.
• You are renumbering or re-sectioning findings from multiple sources into one deliverable that a strict renderer/checker will parse.

Lane output contract (demand this from lanes)

• One ### F<n> — <claim, present tense> [GRADE] block per finding; grade in the closed set {MEASURED, READ, INFERRED, UNVERIFIED} at the END of the claim.
• **Evidence:** line REQUIRED for READ/MEASURED (path:line or URL).
• INFERRED states its reasoning basis; UNVERIFIED states why not checked.
• A ## Still open list and a grade-mix line (N findings: x READ, ...).
• Findings carry Type: (FLAW/GAP/ACCEPTED/INFO) + severity (L×I) when scored.

Getting the full lane reports (Hermes)

The consolidated batch message may arrive late, and the LIVE delegation transcripts (~/.hermes/cache/delegation/live/<delegation-id>/task-N.log) ABBREVIATE long summaries — the final line ends in …(+N chars) and is NOT the full report. Do not merge from the transcripts; you would lift truncated claims.

The authoritative full summaries live in the Hermes state DB (read-only):

sqlite3 "file:$HOME/.hermes/state.db?mode=ro" \
  "SELECT result_json FROM async_delegations WHERE delegation_id='<id>'"

result_json → results[], each carrying summary (may be truncated), summary_full_path (the full markdown, e.g. ~/.hermes/cache/delegation/subagent-summary-N-<ts>.txt), plus model, api_calls, and tokens {input, output}. The live-dir manifest.json holds only paths and status. To record lane tokens in the task tracker without manual counting: task_tracker.py subagent <taskId> --delegation <delegation-id> --description "..." reads the same row. Use the summary_full_path files as the verbatim lift source for assembly.

Assembly steps

1. Extract blocks with ^### (F\d+[ab]?) — (MULTILINE). The id may contain letters (F4b parses); the claim line must end with the grade in brackets.
2. TRUNCATE every block at the first ## header. Lanes embed their own ## Still open + grade-mix sections at the end of their file — lifted verbatim, that header lands INSIDE the last finding's body and ends the renderer's ## Findings capture early. Symptom: the provenance chart counts only the first section's findings (e.g. 13 of 41) while the markdown looks complete. Strip at body.find("\n## Still open") and rstrip separators.
3. Dedupe cross-cutting themes. When two or three lanes independently file the same gap (rotation cadence, missing diagnostics, shared-key auth), consolidate into ONE finding in a dedicated cross-cutting section, with an italic provenance note naming the source lane findings. Never keep both a lane finding and a consolidated twin.
4. ACCEPTED items go to a register section only (never scored, never in the scored sections). Each entry keeps decision citation + residual risk + revisit trigger.
5. Renumber sequentially, then run a stale-cross-ref pass. Renumbering breaks every in-body reference ("see F12", "(F5)", "F2's ..."). Grep the assembled file for every old id pattern and map each to the new id or the register id (E#). A reviewer WILL find these — do the pass before review, not after. This is the most common quality-gate finding on assembled records.
6. Section structure that renderers accept: all ### F# blocks under one ## Findings; #### subheaders for clusters are fine; front sections (## Executive summary, ## Scope & method) and trailing sections (## Still open, ## Verification checklist) must NOT appear between finding blocks. State grade mix + severity counts in the exec summary; ## Still open non-empty.
7. Render as the format gate. A clean render is a real check: the renderer refuses ungraded claims, READ/MEASURED without Evidence lines, and any HTML target inside the repo. Render to a temp dir; never commit the HTML.

Verification checklist

• No ## header inside any finding body (grep for ^## between ### F blocks)
• Old-id grep pass clean (every "see F#/E#" resolves in the final numbering)
• Every READ/MEASURED block has **Evidence:**
• Grade mix + severity counts in exec summary match the rendered chart
• Renderer ran clean to a path OUTSIDE the repo
• ## Still open non-empty, or its emptiness defended

Wave-PR delivery (sequential wave branches under squash merges)

When the work ships as one PR per wave and merges are squash merges, see references/wave-pr-hygiene.md — branch rebasing, draft-gating on owner-side prerequisites, and the gh CLI quirks that look like failures but aren't.

Gotchas

• Trust the renderer's count, not the markdown. A truncated capture still renders "fine" for the part it saw. Check the provenance chart's total.
• Consolidation ≠ summarization. Merge by quoting the lane evidence lines, not by rewriting the claim. The consolidated finding must cite what each source lane evidenced.
• Renumbering is not a mechanical sed. In-body references are prose ("the F5-chain", "F2's zero diagnostic settings") — a plain id-replace pass misses the prose forms. Grep for the ids with context, then read each hit.
• Register ids (E#) are cross-referenced too. Findings reference accepted items; those refs also break under renumbering.