context-routing

name	context-routing
description	Cross-zone retrieval and task-driven bundle resolution for BCOS. /context search ranks docs across every zone in one shared schema; /context bundle (P5) returns a curated, freshness-flagged, source-of-truth-aware context bundle for a declared task. Mechanical-first; LLM is opt-in via --semantic / --resolve-conflicts / --verify-coverage flags only. Invoke with /context.
trigger	/context
version	0.3.0
last_updated	2026-05-11
authority-docs	["docs/_bcos-framework/architecture/context-zones.md","docs/_bcos-framework/architecture/system-design.md","docs/_bcos-framework/architecture/cross-repo-retrieval.md","docs/_planned/wiki-missing-layers/pre-flight-decisions.md","docs/_planned/wiki-missing-layers/implementation-plan.md"]

/context

Cross-zone read foundation for BCOS. Reads the canonical corpus at .claude/quality/context-index.json (built by context_index.py) and exposes:

• /context search — ranked hits across every zone (active, wiki, planned, collections, framework, …). Mechanical BM25; LLM-touching paths require explicit --semantic opt-in.
• /context bundle (P5, not yet implemented) — task-driven curated bundle: by-zone hits, freshness verdicts, source-of-truth conflicts, missing perspectives. LLM-touching paths require explicit --resolve-conflicts / --verify-coverage opt-in.

For zone-scoped sugar, see /wiki search (P2) and /wiki bundle (P5) which share this backend.

Trigger phrases

• /context with subcommands: search, bundle (P5)
• "Find docs about competitor pricing across the whole repo"
• "Search the wiki and the plans for stripe billing"
• "Build me a context bundle for writing a market report" (P5)
• "What do we have on linkedin tone, ranked by source-of-truth?"

Pipeline at a glance

docs/  →  context_index.py  →  .claude/quality/context-index.json
                                              ↓
                       /context search   →   context_search.py
                                              ↓
                                              ranked hits (zone-typed, citation-id'd)

                       /context bundle   →   context_bundle.py  (P5)
                                              ↓
                                              { by-zone, freshness, conflicts,
                                                missing-perspectives, traversal }

Subcommands

Command	Status	Phase	What it does
search <query>	implemented	P2	Mechanical BM25 over the canonical context-index.json; cross-zone or --zone <id> filtered
bundle <profile>	implemented	P5	Task-driven curated bundle per a profile from _context.task-profiles.yml — see bundle.md

Dispatch

1. Identify the subcommand from the first word after /context.
2. Route — read the sibling file in this skill directory and follow its instructions:
   • search → search.md
   • bundle → bundle.md
3. Do not proceed past dispatch before reading the target file.

Guard rails

D-10 strict — mechanical-first, opt-in escalation only

The default path for every subcommand is mechanical Python. No LLM call ever fires implicitly. LLM-touching paths exist only behind explicit flags:

• /context search --semantic — LLM reformulates the query into mechanical candidates.
• /context bundle --resolve-conflicts — LLM resolves unresolved source-of-truth conflicts. (P5)
• /context bundle --verify-coverage — LLM verifies prose-perspective coverage. (P5)

There is no auto-trigger on 0-hit, no auto-fallback, no hidden cost. If the user wants LLM help, they ask for it. If they don't, the engine returns the mechanical answer.

Empty zones are graceful

When a declared zone is absent from the current repo (e.g., _wiki/ not initialized, _collections/ empty), search returns an empty hit list for that zone and surfaces it under zones-skipped-not-present. The agent decides whether to proceed, ask the user, or trigger initialization. Not a hard failure.

Citation IDs are stable

Every hit carries a citation-id of the form zone:slug (for example, wiki:linkedin-tone, planned:wiki-missing-layers). The ID is path-derived and survives whitespace edits. Other agents reference back via this ID.

Sibling-repo hits (when cross-repo retrieval is engaged — see next section) get an additional prefix: <sibling-id>:<zone>:<slug>. Local hits keep the original two-segment form.

Cross-repo retrieval — opt-in, three-stage filter

/context search and /context bundle consult sibling BCOS repos registered with the same umbrella (.bcos-umbrella.json) only when explicitly opted in, and even then only through a three-stage filter:

1. Local search across all zones in this repo (existing behavior).
2. Peek — if local is insufficient, scan sibling metadata (titles, tags, exclusively_owns, headings) for the query. Cheap. No BM25. No body scan. No LLM.
3. Deep-fetch — only when the peek points to a clear sibling winner. Marginal peeks emit a cross-repo-suggestions envelope field but do NOT pull data; the calling agent decides whether to surface to the user.

Flags + config:

• --cross-repo → bypass peek gate, deep-fetch directly (manual override).
• --no-cross-repo → skip cross-repo entirely for this call.
• Default → consult .bcos-umbrella.json.retrieval.auto_fallthrough (ships off; user flips per portfolio).

Hierarchy locked: this-repo first across all zones; cross-repo is last resort. A T5+full-coverage local hit always skips peek entirely — the calling agent doesn't even see a cross-repo-suggestions field.

Framework-default behavior is unchanged: in a repo with no .bcos-umbrella.json or no retrieval block, no cross-repo I/O ever fires. D-10 strict invariant preserved at framework level; per-portfolio opt-in lets a user turn it on for a portfolio of related repos.

Contract: see docs/_bcos-framework/architecture/cross-repo-retrieval.md. The framework consumes a minimum schema from the umbrella's projects.json (id, path, exposes); the umbrella's onboarding skill owns the full write-side schema and writes retrieval.auto_fallthrough: true into each registered sibling by default.

Sibling failures are graceful — cross-repo-status.siblings-skipped[] records each per-sibling problem (timeout, unreachable, malformed-index) without raising. Local result is always returned.

Architectural placement

• context-zones.md — what each zone is and how the path classifier maps to a declarative registry.
• context_index.py — the canonical model that builds context-index.json. Cross-zone retrieval reads this; never re-walks docs/.
• load_zone_registry.py — the loader every cross-zone consumer goes through.
• bcos-wiki/SKILL.md — /wiki search and /wiki bundle are zone-scoped sugar over this skill's backend.

Related skills

• bcos-wiki — wiki zone authoring; provides /wiki search (zone-scoped sugar over this backend).
• context-ingest — entry point for new content; routing decisions read the same zone registry.
• context-audit — CLEAR compliance auditing; reads the same canonical model.
• clear-planner — implementation planning; profile-driven bundles (P5) help the planner ground decisions.

Reference

• search.md — invoke context_search.py; flag handling; output formatting.
• bundle.md — invoke context_bundle.py; profile resolution; conflict surfacing; opt-in LLM escalation gating.