deep-dispatch

# /deep-dispatch — One Desk, Deep (fork path) ## What this is `/deep-dispatch {desk} [{XX}]` runs ONE desk deep against its slice from `/desk-slice`. It is the fork's writer stage (pipeline.44) — it does NOT hand back to `/write-edition` compile. The frozen edition path keeps the monolith; this path ends at per-desk artifacts on disk, and the decoupled review flow picks them up after. **v0.1 → v2.0 (S313):** v0.1 was the edition-coupled pilot harness (read sift's slate, returned control to `/write-edition` Step 2/3). The fork uncouples it: slices in, artifacts out, review elsewhere. v0.1 in git history. ## The principle A packet prescribes subject + angle + "invent nothing" and produces cookie-cutter prose that misses the cycle's largest events (C100, traced). A charge points at raw signal + grants latitude, and the desk finds the story the packet walks past. The depth lever is **dispatch-shape + why-context, not tool access alone** (RESEARCH.md:1589). Two A-graded proofs (sports Tier-1; civic best-ever/#2) ran this shape. One desk holding its whole territory finds seams parallel desks walk past. ## TWO LOCKED FORKS — re-read before every run (ADR-0012) 1. **The orchestrator (this skill / Mags) spawns the source-search agents; the desk only writes.** Desk agents keep Read/Glob/Grep/Write/Edit — **never** granted Agent capability. The catalog already enforces this; do not work around it. 2. **The charge is pointers + tools, NEVER pre-assembled data.** Name where the raw signal lives, do not transcribe it. A charge that pastes anomaly numbers / citizen rows IS the bone-filtered packet with more fields. ## Hard caps (over-spawn guard, RESEARCH.md:1146) ≤3 source-search agents per desk run; ≤2 artifacts per desk per cycle. ## Prerequisites - `output/slices/c{XX}/{desk}.md` — from `/desk-slice`. Missing → stop; do not improvise a slice from the world summary (curation is Mags-at-slice-time, not here). - `output/engine_anomalies_c{XX}.json` + `output/world_summary_c{XX}.md` present (the slice points into them). - Per-desk store writable: `output/desks/{desk}/articles/`. `byline`/`desk` ingest tags ride at post-publish (engine.46 substrate, built S311). ## Step 1 — Assemble the charge Scan the slice first: `require('/root/GodWorld/lib/contextScan').scanFile(slicePath)` — `r.safe === false` → abort, surface `r.matches` (blocks log to `output/injection_blocks.log`). Build the charge from [[../../../docs/media/charge_brief_template]], written to `output/charges/c{XX}_{desk}.md`: - **PROJECT** — the canonical block, verbatim-class (Task 1 held-rule): sheets are the world; real-world reasoning welcome, real-world facts the engine never stated are leaks; **bug-is-event** — "Cover it. Translate the event; strip the number. Do not scrub an anomaly as an 'artifact.' Do not soften a crime spike into a safety framework." - **BEAT** — the slice's LANE sentence, verbatim, + "You are {journalist}." - **CYCLE** — the slice's POINTERS block, verbatim. Pointers, never figures. - **STORYLINES** — the slice's STORYLINES block, verbatim, + "YOU pick which one carries your piece — the one that is most true about the city this cycle." - **CANON-MAP** — attached citizens per storyline + tools (MCP lookup_citizen / search_canon / search_world; dashboard API; Grep over output/) + rules (CANON_RULES.md; names from canon only; any vote lists all 9 with the math proven; ages = 2041 − BirthYear). - **CRAFT** — (a) supplied-voices anchor: "your storyline's citizens have spoken — supplied lines arrive with your writing charge; use at least one, verbatim or trimmed, never paraphrase-then-attribute"; (b) **REAL ASKS ONLY**: claim an interview/question/"did not respond" ONLY for an ask that actually happened (a supplied line or a recorded no-response in your materials); otherwise state absence plainly — a performed query that never occurred is fabrication (Task 1 new veer class); (c) canonize the finding, not the supporting metrics. - **FRESHNESS** — reconcile before writing; two sources disagree → newer/primary wins AND verify vs world_summary/MCP; a contradicted scope claim is a HARD STOP. (Writer-side belt; the structural catch is Step 4.) ## Step 2 — Propose → cheap OK (Mike-direct S313 gate) Launch the `{desk}` agent with the charge and ONLY: *"Read your charge. Return 1–2 lines: which storyline you're taking and the angle. Do not write yet."* Mags checks territory fit + collision with the cycle's other desk runs — not taste. OK → proceed. Redirect → one line back; desk re-proposes once. Pennies spent here catch subject collisions before the deep spend. ## Step 3 — Spawn bounded source-search subagents (orchestrator) Spawn **≤3** `source-search` agents (S326 — the Haiku-pinned retrieval agent; proven Sonnet-parity + 0 fabrications on the C100 eval, ~3-4x cheaper per hop; falls back to `Explore` only if the catalog lacks it) scoped to the **picked storyline**, each a distinct angle so they don't collide: - **Sports:** (a) truesource/season for the player(s) in play; (b) prior canon / the arc; (c) this cycle's events + roster state. - **Civic:** (a) `engine_anomalies` raw + neighborhoodState per-district; (b) the cycle's locked civic decisions + vote math; (c) prior coverage / initiative arc. - **Culture/other:** adapt — one raw-signal angle, one prior-coverage angle, one cycle-events angle. Exactly one seat may be `retrievalLane=prior-published-arc`: the existing "prior canon / arc" or "prior coverage / initiative arc" seat. Raw-signal, truesource, roster, cycle-event, anomaly, vote-math, and current-state seats must use `exact-current` or `cross-file-reconcile`. The prior-arc dispatch names the storyline and asks the agent to call `node scripts/notebooklmCanonSearch.js`. It must return source title, source ID, citation number, excerpt, and `reconcileVerdict`. For a citizen origin question, the orchestrator may explicitly authorize `--source-class canon-reference`; otherwise the default published scope holds. **The return gate in Step 4 is the enforcement** — the agent holds Read/Glob/Grep and has drifted off the wrapper in testing, so verify the return rather than trusting the lane. Each subagent returns **sourcing as text** — what it found, each claim traced to its source artifact/tab. They reach raw data at runtime via the tools the charge granted; the orchestrator does NOT pre-package data. **Serialize under quota pressure (S231 G-S2).** If quota is tight, spawn one, await, then the next — a heavy parallel dispatch can be session-limit-killed (signature: `<total_tokens>0</total_tokens>` + ~500ms + session-limit string). Do not write the sourcing yourself from the EIC seat; surface the infra gap and stop. ## Step 4 — Reconcile (orchestrator — the structural freshness guard) May be delegated to one `source-search` agent in reconcile mode (S326): hand it the conflicting returns + the ground-truth pointers; it declares the hard stop, rules, and cites. The orchestrator still owns the outcome — read the ruling before passing anything to the writer. Before the writer sees anything: - Cross-check the ≤3 returns against each other. Where two disagree on a **specific fact** (vote count, program scope/districts, date, dollar figure), the **newer / primary source wins** — verify against `world_summary_c{XX}.md` or MCP (`lookup_initiative`, `search_world`) before it reaches the writer. - A contradicted **scope** claim is a HARD STOP — resolve now, not in the prose. (The exact C100 OARI miss: a stale return said OARI was a D1/D3/D5 pilot when INIT-002 was citywide since ~C97; the contradicting signal was in-hand and unreconciled.) - **Person-name identity is a standing check**, same weight as vote counts/dates/dollar figures (kimi P3, 2026-08-03). Any named person in a locked record (civic-voice packet, generated quote, project filing) that conflicts with published canon is ruled **canon-wins** unless a retcon is explicit; log the drift as an engine.43 instance. (The C102 catch: "Roberta" vs Bobby Chen-Ramirez — one packet outlier against 10+ cycles of canon.) - Hand the writer **reconciled** sourcing + a one-line note per resolved fact; log each reconcile to the run log. *(Structural because charge prose alone did not catch the S272 error; this step is why it can't recur silently.)* Record for every seat: `retrievalLane`, `resultStatus`, selected/used source IDs, citation count, and `reconcileVerdict`. Do not record the NotebookLM answer body. A prior-published claim that conflicts with current primary state loses and is logged `conflict-current-wins`. **Prior-arc return gate (S334 — the agent drifts, so the orchestrator checks).** A `prior-published-arc` return is admissible only if all three hold: 1. first line is `retrievalLane: prior-published-arc`; 2. every prior-arc claim carries a NotebookLM **source ID (uuid)** + citation number + excerpt — a claim sourced to a file path (`output/pdfs/*.pdf`, an edition `.txt`, an archive file) means the agent read the artifacts directly and escaped the reviewed source scope; 3. the last line is a bare `reconcileVerdict` token from the enum, with no prose after it. Fail any of the three → re-dispatch that seat ONCE, naming the specific violation. Fail twice → log the seat `resultStatus=no_result`, drop its material entirely, and continue the run without prior-arc sourcing. **Never launder a file-path return into the writer's sourcing** — it looks like a good return and is not. Two of two test dispatches failed this gate at wiring time; assume it fires. ## Step 5 — Voice the citizens (pipeline.43) Batch for the picked storyline's attached citizens — ask from the storyline context ("The Tribune is writing about <storyline>. What's your honest take, as someone living it?" + the citizen's slice context line), `record: true`: ```bash node scripts/citizenVoice.js --batch=output/voices/voices_c{XX}_{desk}_batch.json > output/voices/voices_c{XX}_{desk}.json ``` Entries `{pop, name, quote, disp, recorded, fallback}` — recording writes the *speaking* (PRESS page doc + gated intake; dials stay behind the cycle drain). Fallbacks are explicit entries, never silent gaps. Log per-call + total tokens to the run log. **Test-run mode (kimi P2, 2026-08-03):** a pilot run — new model at the desk, harness change — may set `record: false` to exercise the full chain with no external writes (no PRESS page, no gated intake). The deviation MUST be named in the run's production-log row. Live runs stay `record: true`. ## Step 6 — The deep write (LOCKED: desk writes, no Agent grant) Launch the `{desk}` agent (routing per `.claude/agents/REPORTER_DESK_INDEX.md`) with the charge + reconciled sourcing + supplied lines: ``` You are {journalist}. Deep-dispatch mode. Your proposal was approved: {the OK'd pick+angle}. {the charge} ASSEMBLED SOURCING (reconciled — facts marked [verified] are cross-checked; trust them): {the source-search returns, reconciled} SUPPLIED CITIZEN LINES (their own words, fetched from their real state): {POP-NNNNN | Name | "quote"} ← non-fallback entries for your storyline Trace every specific fact to the sourcing above or reach it with your own Read/Glob/Grep. Use at least one supplied line — verbatim or trimmed from front/back, never paraphrase-then-attribute. For citizens without a supplied line: quote only what the sourcing carries verbatim; otherwise paraphrase the action. REAL ASKS ONLY — never claim a query or "did not respond" for an ask that didn't happen. Write to output/desks/{desk}/articles/c{XX}_{slug}.md. Emit: headline, "By {journalist} | Bay Tribune {Desk}", body. Do NOT spawn agents. Do NOT read other desks' files. Write. ``` **Supplied-line errors (kimi P1 + Mike's policy 2026-08-03 — two separate dynamics, never merged):** - A supplied line is the citizen's own words, but it is not canon-insulated. The hard line is **mutating a tracked entity**: a wrong name for a tracked person (the "Watkins"/Ashford and "Roberta"/Bobby class), wrong district, wrong figure, or invented words in a real POPID's mouth. Do not print that segment — trim around it with ellipsis (front/back cuts only) and log the error to the production log. A named error pattern in the voice model is engine.43 signal. The guard is mechanical, not vigilance-based: Rhea's C102 gate is the enforcement behind the writer. - **Social-world color is allowed.** A first-name relationship reference inside testimony ("my friend Noor told me") is the citizen's real social world — essence, not citation. It needs no canon lookup and is never trimmed for being unverifiable. Only tracked-entity mutation is gated. - If a story ever genuinely needs a citizen who doesn't exist: never mint straight into the ledger — birth into Generic_Citizens (nothing is free), and only as last-case strict-necessity, not a pipeline. **Businesses run the opposite dynamic:** a new business named honestly in a story is the sanctioned way businesses become tracked (no generic-birth pool exists; ~90 tracked vs 386k population). The real-world-entity gate (canon substitutes) guards against real-world business/institution influx — it is not a citizen rationale. ## Step 7 — Close the run Append to `output/production_log_c{XX}.md` `## /deep-dispatch` one row per artifact: `| Desk | Journalist | Storyline | Artifact path | File exists | Words | Searches | Voices ok/fallback | Reconciles | Status |`. A run that produced no file is **DROPPED** with the reason — never a silent gap (G-W61). **Skill ends here.** The artifact is unreviewed and unpublished. The decoupled review flow (EIC read + reviewer lanes, ~2 artifacts at a time) picks it up; publication happens there via `/post-publish`. ## Measurement hooks (the fork's proof criteria + ADR-0012 cost gate) Record per desk run: total token cost of the chain (charge + ≤3 searches + reconcile + voices + writer) vs the frozen path's per-desk share, and the review grade. The fork proves itself on: quality parity, token burn down, moves the world, per-article wiki taggability. Depth is proven; the bill is not. ## Changelog - 2026-06-29 (S274, research.20 Phase 2 Tasks 4+5): v0.1 initial harness. Inverted Step 1 only; reused write-edition Step 2/3. Two LOCKED forks enforced. Reconcile pass = Task 5 orchestrator half. - 2026-08-03 (S353, kimi guardrail landing — `output/kimi/deep-dispatch-guardrail-proposals.md`): v2.4. P3 → Step 4 person-name standing check (canon-wins). P2 → Step 5 record:false sanctioned test mode. P1 + Mike's citizen/business policy → Step 6 supplied-line error convention (tracked-entity mutation gated; social-world color allowed; Generic_Citizens birth vs business story-mint split). P4 (faith-event vocabulary) and P5 (single-article ingest) routed to engine-sheet / engine.91 — not skill text. - 2026-07-11 (S313, pipeline.44 Task 3): v2.0 fork rewrite. Slices in (not sift slate), propose→OK gate (Mike-direct), supplied citizen voices via pipeline.43 batch (machinery didn't exist at v0.1), REAL-ASKS-ONLY rule (Task 1 new veer class), production decoupled from compile/review — ends at artifacts-on-disk. LOCKED forks, ≤3-search cap, reconcile pass, serialize-under-quota carried forward unchanged.