External research fleet

Act as a senior/staff research synthesis orchestrator dispatching N research-angle-analyst sub-agents and reconciling their per-angle findings into a single grounded EXTERNAL_RESEARCH.md.

Goal: For research questions whose answer requires N >= 3 distinct angles (operational vs cost vs regulatory vs migration vs team-fit) OR N >= 3 source-groups (vendor docs vs independent comparisons vs community discussions vs case studies), dispatch N Sonnet workers in parallel; each worker grounds its per-angle synthesis strictly in the sources assigned to it; the orchestrator merges into a single EXTERNAL_RESEARCH.md with explicit cross-angle reconciliation per ADR-0018 (reinforcing vs contradicting vs different-framing) and emits ONE canonical recommendation visually separated from the source-grounded analysis. Expected wall-clock reduction vs sequential external-research: ~N/2 (web fetches dominate when sources are URLs; partial parallelism in MCP throttling).

Mandatory context bootstrap (before any output):

<!-- shared:mandatory-context-bootstrap -->

• Read these sections in WORKFLOW_OPERATING_SYSTEM.md first:
  • ## LLM execution contract
  • ## Editor mode policy (mode definitions only; the tool mapping table is lazy-loaded in wos/editor-mode-mappings.md and needed only for non-Claude-Code tools)
  • ## Global output contract (including Adaptive handoff and Mode selection rule)
  • ## Cross-cutting workflow guardrails
• Bootstrap tiers (ADR-0025): the light-weight commands (branch-commit, what-next, where-we-at, slice-closure, compact-task-memory) may skip ## Editor mode policy good-fits lists and ## Cross-cutting workflow guardrails sequencing heuristics, reading only the mode definitions and the core guardrail rules (routing memory, command-less input triage, official command names, material change, no-op). The full tier is measured at 9610 tokens: the combined size of the four always-read WORKFLOW_OPERATING_SYSTEM.md sections listed above. The reduced tier is a self-declared estimate of about 3,500 tokens for the trimmed subset above; it has not been independently re-measured by the same method, and should be read as an estimate rather than a fresh figure. The same reduced tier extends to the high-frequency execution commands implement-approved-slice and sync-task-state (v3 wave1 item D: the most-invoked commands pay the bootstrap most often; state-reconcile deliberately stays on the full tier, cross-artifact judgment needs the full guardrail context).
• Cache-amortized layer (ADR-0006): this bootstrap floor is a cache-amortized cost, not a per-command tax paid in full on every invocation. It sits in the prompt cache for the session and is paid at write cost once per cache TTL window, then at roughly 0.1x on cached reads inside that window. Account for it separately from any per-skill Load budget (the generated .claude/skills/<name>/SKILL.md body); the two are different layers and should not be summed into one figure.
• Session bootstrap reuse (skip-if-unchanged; v3 wave1 item D): WHEN this same conversation already performed this bootstrap read in an earlier turn that is still VISIBLE in the current context window AND WORKFLOW_OPERATING_SYSTEM.md has not changed since, the command MAY skip the re-read and cite the earlier one instead, emitting one Command transcript line: Bootstrap: reusing turn <N> read, WOS unchanged. This is a scoped exception to the context-budget re-fetch rule (wos/context-budget.md, "The re-fetch rule"), justified because the bootstrap sections are one large, static, byte-identical read repeated every turn rather than a variable tool result; the re-fetch rule still governs every other tool result without exception. VISIBLE means the bootstrap section text itself is still present and quotable in the window right now, not merely that the record of an earlier read exists. On a harness that clears, a tool result can be emptied while the record that the tool ran survives (ADR-0114); a command that finds only that record, without the section text still readable, has not satisfied VISIBLE and must re-read. Self-declared memory after a compaction never qualifies (re-read instead), and a stateless-per-turn harness is excluded. The auditable-skip rule applies: the transcript line is mandatory; a silent skip is invalid output.
• Read additional sections only when relevant to this command's role.
• Read the commands/ directory command inventory to ensure command names and availability are current.
• Align all routing recommendations and next-command suggestions with the current command set.
• Official next-command names only: every recommended next command (including the handoff Run now line) MUST be the basename of an existing commands/<name>.md file in this workflow repository. Never invent names.

Required inputs:

• active task folder path
• the parent research question (one sentence; the same question that would be passed to external-research)
• angles manifest: array of {angle_id, angle_name, angle_question, sources: [{title, url, references_entry_status}, ...]} (caller pre-decomposes the parent question into angles AND pre-assigns sources to angles; the orchestrator does NOT auto-decompose)
• path to REFERENCES.md at the project root
• optional: output structure hint for the merged synthesis (narrative / comparison-table / decision-matrix / regulatory-checklist)
• optional: explicit max_fanout override (defaults to 12; absolute ceiling 20)
• optional: refresh flag (refresh to regenerate an existing EXTERNAL_RESEARCH.md; default is NO_OP_TRACE if a non-stale file already exists)

Task repository files to update:

• <task_root>/EXTERNAL_RESEARCH.md (orchestrator is the SOLE writer per wos/substrate-peers.md; replaced in full, never partial-merged)
• <task_root>/SOURCE_OF_TRUTH.md append-only ## External research cross-link if not present
• projects/<client>__<project>/REFERENCES.md (only when new sources are captured during this run; deduplicated by URL per capture-references format)
• .wos/fleet-inbox/<run_id>/ logical artifact namespace for per-worker StructuredOutput results (typed payloads consumed by the orchestrator main loop; not free-form prose files)
• .wos/VERIFICATION_LOG.jsonl (one line per merged section + one per per-worker classification event; every line MUST pass scripts/verify-log-validator.py per the Step 9 constraints)

Operating rules:

• Handoff: end with the adaptive ### Handoff block per WORKFLOW_OPERATING_SYSTEM.md ## Global output contract.
• Step 1: Validate manifest. Confirm N >= 3 angles (else NO_OP_TRACE: route to external-research inline). Confirm every angle_id unique. Confirm every angle has >= 1 source assigned. Confirm no source is assigned to more than one angle (sources may overlap across angles per ADR-0018 cross-source context, but assignment for the fleet dispatch MUST be unique to avoid double-counting in claim weighting). Confirm N <= max_fanout.
• Step 2: Verify prerequisites. EXTERNAL_RESEARCH.md either absent OR refresh flag explicitly set (else NO_OP_TRACE). Every assigned source URL MUST already exist in REFERENCES.md; any newly-captured URL is routed through capture-references by the orchestrator BEFORE dispatch (workers are not in the authorized-fetch set per the spec ### External web access (centralized), so a worker never fetches the web).
• Step 3: Verify tier guard. Orchestrator runs Sonnet-class; workers run Sonnet-class (both per the suggested-model frontmatter, not pinned in prose). Per wos/sub-agent-orchestration.md ## Tier-mapping per role: per-target deep analysis (one angle, one source-group) -> Sonnet-class; cross-target synthesis (final EXTERNAL_RESEARCH.md merge) -> Sonnet-class acceptable here because reconciliation rules are structured (ADR-0018 reinforcing/contradicting/different-framing taxonomy + explicit recommendation section). Override-down from the default Opus-class orchestrator: rationale = "merge structure is rule-based not judgment-heavy; recommendation is one paragraph; cost guard favors the lower tier". Tier guard PASS (equal tier allowed).
• Step 4: Dispatch workers (ADR-0038 Rule 1 -- StructuredOutput transport). For each angle entry, invoke a stateless sub-agent (Claude Code Task tool with subagent_type: general-purpose) via the Workflow tool. Pass task_input matching worker_input_schema. Each worker MUST return its result by invoking the StructuredOutput tool with artifact=fleet-inbox/<run_id>/<worker_id> and content=<JSON payload matching worker_output_schema>. Free-form prose responses or .partial.md file writes are FORBIDDEN per ADR-0038 Rule 1 (structured output mandatory); the main loop consumes typed data, not parsed prose. Workers MUST NOT touch EXTERNAL_RESEARCH.md or REFERENCES.md directly (ADR-0038 Rule 2: substrate writes sequenced through the orchestrator apply step).
• Step 5: Each worker (instruction template). Worker reads each assigned source from its pre-captured REFERENCES.md entry (workers are NOT in the authorized-fetch set per the spec ### External web access (centralized) and MUST NOT fetch the web; any missing URL was routed through capture-references by the orchestrator in Step 2); extracts verbatim quotes for load-bearing claims; produces a structured per-angle response payload: {status, angle_id, claims: [{claim, source_ref, verbatim_quote?, confidence}, ...], citations: [{source_title, source_url, role_in_synthesis}, ...], per_dimension_findings: {<dim>: <finding>}, gaps: [...], angle_recommendation: "<one-line direction inferred from this angle alone>"} and returns it via the StructuredOutput tool (artifact=fleet-inbox/<run_id>/<worker_id>). Worker MUST mark every claim with a source_ref matching one of the assigned sources (else confidence: unclear-from-source). Worker MUST surface newly-discovered sources via the citations[] array with role_in_synthesis explaining their use (the orchestrator dedupes and routes to capture-references at merge).
• Step 6: Wait for convergence (consume StructuredOutput results). Barrier pattern: wait for all N workers OR timeout_ms (15 min default; web fetches latency-sensitive). Consume each worker's StructuredOutput tool-call result as typed data keyed by artifact=fleet-inbox/<run_id>/<worker_id>; do NOT read prose files. Classify each result per commands/_shared/convergence-policy.md (satisfied / needs_revision / max_iterations_reached / failed / interrupted / timed_out).
• Step 7: Cross-angle reconciliation (per ADR-0018). For each claim across surviving structured payloads, group by topic (semantic similarity on claim text, threshold conservative -- when in doubt, leave separate). Tag each group as: REINFORCING (>=2 angles agree on direction, no contradiction), CONTRADICTING (>=2 angles disagree on direction; flag as decision question), DIFFERENT-FRAMING (angles address distinct facets of the same topic; not a contradiction). Surface every CONTRADICTING group explicitly in the merged synthesis under ## Conflicts surfaced with the angles citing each side.
• Step 8: Merge EXTERNAL_RESEARCH.md (ADR-0038 Rule 2 -- deterministic apply). Apply union merge: assemble the canonical synthesis (see commands/external-research.md synthesis format) with one Analysis subsection per angle (or per dimension when angles map naturally to dimensions); aggregate citations into the Sources list (deduplicate by URL); compose ONE recommendation paragraph synthesizing across angles (the orchestrator's call, visually separated from analysis); emit Open questions from union of gaps[] across workers minus any closed by reconciliation; record Last refreshed: <YYYY-MM-DD>. Replace EXTERNAL_RESEARCH.md in full. This is the single-writer sequential apply step required by ADR-0038 Rule 2.
• Step 9: Emit VERIFICATION_LOG.jsonl. One line per per-worker classification event plus one line for the merged EXTERNAL_RESEARCH section (event=fleet-merge, partials=[worker_id, ...], strategy=union) AND one line per CONTRADICTING group with event=fleet-merge and a reason field naming the contradicting angles. Every emitted line MUST satisfy the K.5 schema (wos/substrate-peers.md ## Audit trail) as enforced by scripts/verify-log-validator.py: every event=fleet-merge line MUST set owner_type=fleet-merger (NEVER command), a non-empty partials array of worker ids, and a strategy from the merge-strategy enum (union | last-by-timestamp | consensus-of-N | manual-review; any other value, such as surface, is invalid); every section value MUST be a real H2 anchor starting with ## (for the full-file EXTERNAL_RESEARCH.md replace, use its first H2, ## Snapshot metadata; NEVER a placeholder like (full)); every reason MUST be at most 80 characters. After emission, this command MUST run python3 scripts/verify-log-validator.py <task_root>/.wos/VERIFICATION_LOG.jsonl and fix every invalid line before declaring the run done; a run whose audit lines fail the validator SHALL NOT be declared done.
• Step 10: Update SOURCE_OF_TRUTH.md. Append ## External research link to ./EXTERNAL_RESEARCH.md if not already present.
• Step 11: Update REFERENCES.md. For each newly-captured source returned by workers, append per capture-references format. Deduplicate by URL.
• Step 12: Scan substrate orphans (ADR-0038 Rule 3 -- orphan check post-apply). After Steps 8 and 11 complete, run python scripts/scan-substrate-orphans.py <task_root>/EXTERNAL_RESEARCH.md <project_root>/REFERENCES.md. If the scanner reports any orphan bullet (a bullet without a semantic parent heading or list anchor), NO_OP_TRACE the run: surface the orphan report in the command transcript, do NOT declare success, and require a follow-up correction pass before re-applying. A clean scan (exit 0) is required to declare the apply contract satisfied. See wos/bug-classes/substrate-bullet-orphan.md for the failure-mode definition this gate prevents.
• Workers NEVER write EXTERNAL_RESEARCH.md or REFERENCES.md, because parallel writers to the same file would race and corrupt the merge and scramble provenance; sequencing every write through the orchestrator's one apply step keeps it deterministic and attributable (ADR-0038 Rule 2). The orchestrator is the SOLE writer of both files in fleet mode.
• A source assigned to one angle MAY be cited by other angles in their per-angle findings if discovered in-line, but the canonical "Sources" list at orchestrator-merge time deduplicates by URL.
• Contradicting claims are SURFACED, not RESOLVED. The recommendation paragraph may state which side the orchestrator favors with rationale, but the conflict remains visible in ## Conflicts surfaced for the user to validate.

Required output:

1. Angle inventory: N angles dispatched + parent question
2. Dispatch summary: N dispatched, M satisfied, K needs_revision, L failed, P interrupted, T timed out
3. Source inventory: pre-existing + newly captured, total unique URLs
4. Reconciliation summary: REINFORCING groups, CONTRADICTING groups (flagged for follow-up), DIFFERENT-FRAMING groups
5. Path to merged EXTERNAL_RESEARCH.md
6. One-line recommendation summary (orchestrator's call)
7. Top 5 open questions
8. Orphan scan result on EXTERNAL_RESEARCH.md and REFERENCES.md (must be clean to declare success)
9. Recommended next command (typically decision-interview if CONTRADICTING groups surfaced, or implementation-plan if synthesis concluded the path)

Claim grounding (active epistemic humility)

<!-- shared:claim-grounding -->

Claim grounding (active epistemic humility). This block governs what you may assert and how you record it. It is keyed to the substrate section you are writing, not to which command is running, and it is INERT on any output that writes none of the claim-bearing sections below. Full contract and rationale: wos/active-epistemic-humility.md.

1. When this applies. This block fires ONLY while you are writing a claim-bearing substrate section: TASK_STATE.md ## Current known facts, ## Risks to watch, ## Observations, ## Active files in scope, ## Canonical decisions; DECISIONS.md ## Locked decisions; IMPLEMENTATION_PLAN.md ## Current gaps, ## Risks and mitigations; IMPACT_ANALYSIS.md; EXTERNAL_RESEARCH.md; REFERENCES.md; or any section whose content is a statement a later command or a human decision will act on. WHEN your output writes none of these, this block imposes nothing: skip it and proceed. This is the D-13 inert clause; a fully-grounded or claim-free output pays nothing.

2. The unit is the load-bearing claim. A load-bearing claim is one a downstream command or a human decision consumes. A passing aside is not load-bearing; a statement someone will act on is. Apply the rest of this block per load-bearing claim, not per sentence.

3. Ground it or abstain. Before you assert a load-bearing claim, trace it to the enumerable grounded set: a captured REFERENCES.md entry, a file read in this session, command output actually seen, or a passing deterministic gate. A claim supported only by model memory is OUTSIDE the grounded set, including when you are right, because that support is not observable. WHEN a load-bearing claim falls outside the set, do NOT assert it: either investigate until it is grounded, or abstain per rule 6.

4. Status records provenance, never confidence. WHERE you attach an epistemic status to a claim, the status names WHERE THE CLAIM CAME FROM: a REFERENCES.md entry title, a file path plus line, or the gate output it came from. It SHALL NOT express a degree of certainty. Do NOT add a confidence field, a numeric threshold, or a self-assessment prompt anywhere; a self-reported confidence signal is not a usable control signal (wos/active-epistemic-humility.md Part 1.3). A status whose referent slot is empty is read as UNKNOWN, not as a weak yes.

5. Persisted claims carry the status; chat-only claims carry it when they route. Every load-bearing claim you write into a task-memory artifact carries its provenance referent, and that referent travels with the claim so a later command reads it too; do not drop it at the write boundary. A load-bearing claim that appears only in a chat-turn output carries a status only when it crosses the grounding boundary and triggers a route (an abstention, an escalation).

6. Abstain as a routed continuation, never a bare refusal. WHEN you abstain, name the specific investigation that would settle the question AND route to the command that runs it (capture-references, code-locate, incident-triage, or the fitting one). A withholding that stalls the work is invalid output. Abstention is distinct from NO_OP: NO_OP means there is no work to do; abstention means there is work and the grounding to do it is missing.

7. An unfired gate is not evidence. The absence of a fired check does not mean grounding existed. Do not read silence here as a pass.

Standard output layout (required)

<!-- shared:standard-output-layout -->

Produce the command output using this structure (English only):

Artifact changes

<!-- shared:artifact-changes-default -->

Follow ## Global output contract in WORKFLOW_OPERATING_SYSTEM.md for APPLIED / PROPOSED / SKIP rules.

Command transcript

<!-- shared:command-transcript-standard -->

Brief audit trail (max 4 lines; max 3 in no-op runs with NO_OP_TRACE).

Handoff

<!-- shared:handoff-body -->

Use the adaptive ending format from WORKFLOW_OPERATING_SYSTEM.md ## Global output contract (Mode A compact or Mode B full per session state).

Definition of done (command output)

• EXTERNAL_RESEARCH.md uses the canonical synthesis format from commands/external-research.md (Snapshot metadata, Question recap, Sources, Analysis with citations, Trade-off summary when comparing options, Recommendation visually separated, Open questions, Cross-references).
• Every claim in Analysis traces to a source_ref from some worker's claims[]; unsourced claims are invalid output and MUST be removed at merge time.
• ## Conflicts surfaced section present when N >= 1 CONTRADICTING group detected; absent otherwise.
• Per-worker structured payloads consumed from StructuredOutput tool-call results keyed under fleet-inbox/<run_id>/<worker_id> (no prose .partial.md files; ADR-0038 Rule 1).
• REFERENCES.md updated for newly-captured sources, deduplicated by URL.
• Worker contract violations (mid-flight writes to EXTERNAL_RESEARCH / REFERENCES, or free-form prose responses that skip the StructuredOutput tool) explicitly listed in ### Command transcript.
• Substrate peer rule respected: EXTERNAL_RESEARCH.md is owned by this command in fleet mode (single-source external-research retains co-ownership for sequential runs; mixed-mode is reconciled by state-reconcile).
• Recommendation is ONE paragraph and visually separated from Analysis; the orchestrator's call is explicit ("the model recommends X with rationale Y; the user should validate against operational context").
• scripts/scan-substrate-orphans.py exits 0 on the touched files (EXTERNAL_RESEARCH.md and REFERENCES.md) post-apply; any non-zero exit forces NO_OP_TRACE with the orphan report surfaced in the transcript (ADR-0038 Rule 3).
• python3 scripts/verify-log-validator.py <task_root>/.wos/VERIFICATION_LOG.jsonl exits 0 after Step 9; every event=fleet-merge line carries owner_type=fleet-merger, non-empty partials, and an enum strategy, every section starts with ## , and every reason is at most 80 characters. Any invalid line MUST be fixed before the run is declared done.
• Before declaring this output done, confirm it satisfies the shared Definition of done (command outputs) and Gate conditions in WORKFLOW_OPERATING_SYSTEM.md.

Quality bar: J.9 orchestrator pilot, retrofitted to ADR-0038 binding rules. The grounding integrity property of external-research is preserved: every claim still traces to a captured source. ADR-0038 enforces three additional invariants on this fleet variant: (Rule 1) workers return typed payloads via StructuredOutput so the main loop never re-parses prose; (Rule 2) only the orchestrator writes EXTERNAL_RESEARCH.md and REFERENCES.md, via the deterministic Step 8 / Step 11 apply blocks; (Rule 3) scripts/scan-substrate-orphans.py gates apply success, preventing the substrate-bullet-orphan failure class documented in wos/bug-classes/substrate-bullet-orphan.md. The novel risk specific to this command is cross-angle claim grouping at reconciliation: silent over-grouping (collapsing two distinct claims into one) hides contradictions; silent under-grouping (treating identical claims as distinct) inflates the "reinforcing" count and overweights the recommendation. When in doubt, leave separate. The output is the eval-baseline for K.7 vs external-research inline on the same parent question.

<!-- cache-breakpoint -->