business-context-os

LLM wiki kit

First committed by walm00

Published in walm00/business-context-os. Publisher page

22 GitHub starsUpdated 2026-05-15

What it is

CLEAR Context OS is the maintenance layer your second brain never had: ownership boundaries enforced mechanically, staleness surfaced on a schedule, drift propagated across linked docs, a dispatcher that learns from your clicks. One source of truth per topic. Searchable. Task-aware.

— the description by walm00

AutoGineer overview, from the brain’s own files

CLEAR Context OS is a framework designed to solve the problem of silent context degradation in AI agents. By implementing a structured methodology called CLEAR (Contextual Ownership, Linking, Elimination, Alignment, Refinement), it provides a maintenance layer that manages knowledge through automated scheduling, ownership enforcement, and drift detection. It organizes information into specific zones—such as canonical data, wiki-based derivatives, and verbatim evidence—ensuring that agents always operate on the most accurate, source-of-truth-ranked information. The system includes a dispatcher for scheduled maintenance, a local dashboard for monitoring, and integration with Claude Code to facilitate onboarding and task-aware context retrieval.

How it works

Memory model: LLM wiki

AutoGineer overview, from the brain’s own files

The system organizes information into a folder hierarchy where the path itself acts as a trust signal for the LLM. It uses a dispatcher to run scheduled maintenance jobs that check for staleness, propagate updates across linked documents, and audit content health. New information is ingested through a routing process that places content into specific zones: canonical data for facts, wiki pages for operational knowledge, and collections for legal evidence. The system learns from user interactions to suggest fixes and uses frontmatter metadata to allow agents to filter and retrieve information efficiently without full-text scanning.

Folders
Folder	Role	Files
.claude	agent-config	193
.github	agent-config	6
docs	docs	68

The agent reads

docs/_bcos-framework/guides/folder-conventions.md evidence: CLAUDE.md
docs/_inbox/ evidence: README.md

The agent writes

docs/_inbox/ evidence: CLAUDE.md

Requirements

Works with: Claude Code, OpenCode

Set it up

This brain runs code automatically: 5 hooks (claude-code), 1 script. Review .claude/settings.json, install.sh before opening the folder in Claude Code. No step below turns a hook on for you.
- .claude/settings.json: bash "$CLAUDE_PROJECT_DIR/.claude/hooks/auto_save_session.sh"
- .claude/settings.json: bash "$CLAUDE_PROJECT_DIR/.claude/hooks/precompact_save.sh"
- .claude/settings.json: "$CLAUDE_PROJECT_DIR/.claude/bin/python3" "$CLAUDE_PROJECT_DIR/.claude/hooks/post_edit_frontmatter_check.py"
- .claude/settings.json: "$CLAUDE_PROJECT_DIR/.claude/bin/python3" "$CLAUDE_PROJECT_DIR/.claude/hooks/post_edit_frontmatter_check.py"
- .claude/settings.json: "$CLAUDE_PROJECT_DIR/.claude/bin/python3" "$CLAUDE_PROJECT_DIR/.claude/hooks/post_commit_context_check.py"
- install.sh: sh install.sh

Clone the repository at the commit AutoGineer scanned:

git clone https://github.com/walm00/business-context-os.git && cd business-context-os && git checkout 463568372f31f24fa6c01121c38f989845ca7c84

Claude Code reads CLAUDE.md from this folder automatically.
cd business-context-os && claude
AutoGineer scanned these files at commit 4635683.
First committed by walm00. Published in walm00/business-context-os. Licence: MIT License.

Usage

AutoGineer overview, from the brain’s own files

Automating the maintenance of project documentation and SOPs.
Providing AI agents with a reliable, freshness-aware knowledge base.
Managing ownership boundaries across large, multi-topic documentation sets.
Syncing and auditing context across linked documents to prevent drift.
Generating daily digests of system health and pending maintenance tasks.

“Install BCOS from https://github.com/walm00/business-context-os and walk me through onboarding.”: Claude clones the repository, sets up the environment, and guides the user through an interactive onboarding process to configure the context architecture.
“/context bundle market-report:write”: The agent retrieves a curated, freshness-aware package of context specifically relevant to writing a market report.
“Run the audit twice a week.”: The dispatcher updates the maintenance schedule to perform an audit of the knowledge base on a bi-weekly cadence.
“/wiki run https://stripe.com/docs/api”: The system fetches the URL, summarizes the content, and creates a cited wiki entry within the knowledge base.

Documentation by walm00

<h1 align="center">CLEAR Context OS</h1> Knowledge that maintains itself — so your AI doesn't act on what was true last month. <a href="#install">Install</a>  ·  <a href="#what-it-does">What it does</a>  ·  <a href="#the-wiki-zone">Wiki zone</a>  ·  <a href="#what-ships">What ships</a>  ·  <a href="docs/_bcos-framework/guides/getting-started.md">Getting Started</a> <img src="https://img.shields.io/badge/claude_code-ready-blueviolet" alt="Claude Code"> <img src="https://img.shields.io/badge/methodology-CLEAR-green" alt="CLEAR Methodology"> <img src="https://img.shields.io/badge/install-one_command-success" alt="One-command install"> <img src="https://img.shields.io/badge/skills-17-blue" alt="17 skills"> <img src="https://img.shields.io/badge/maintenance_jobs-12-orange" alt="12 maintenance jobs"> <img src="https://img.shields.io/badge/license-MIT-lightgrey" alt="MIT">

Your AI had perfect context — two weeks ago. Then your strategy shifted, three meetings redefined priorities, and the docs you wrote on Monday are wrong by Friday. CLEAR Context OS is the maintenance layer your second brain never had: ownership boundaries enforced mechanically, staleness surfaced on a schedule, drift propagated across linked docs, a dispatcher that learns from your clicks. One source of truth per topic. Searchable. Task-aware. Never silently stale.

Install

Easiest — let Claude do it. Open Claude Code in any directory and paste:

"Install BCOS from https://github.com/walm00/business-context-os and walk me through onboarding."

Claude clones the repo, picks the right install path for your platform (Windows/macOS/Linux), runs setup, and steps straight into onboarding. No prereq checks to debug — if something's missing, Claude resolves it interactively.

Or manually, two commands:

curl -fsSL https://raw.githubusercontent.com/walm00/business-context-os/main/install.sh -o /tmp/bcos-install.sh
bash /tmp/bcos-install.sh

(Requires bash, curl, and tar on PATH — fine on macOS/Linux and on Windows under WSL or Git Bash. Two steps instead of a single bash <(curl …) pipe so Claude Code's permission gate auto-approves both commands — you don't need to drop out to a terminal.)

Drop anything you already have — SOPs, brand docs, meeting notes, exports — into docs/_inbox/ (skip if you're starting fresh). Then open Claude Code in that repo and say:

"Launch BCOS onboarding."

That's the only prompt you need. Onboarding inspects your inbox, scans the rest of the repo, asks one Q&A round, and proposes the architecture for review. Works the same whether the repo is fresh or already full of context.

From zero to running

  INSTALL  →  DISCOVER  →  ALIGN  →  SCHEDULE  →  RUNNING

#	Step	What happens
1	Install	One bash line drops the framework into any repo
2	Discover	Claude scans your repo + any MCP you've connected (Drive, Notion, Slack, Gmail, your CRM — whatever's wired in) for existing context, plus anything you've dropped into `docs/_inbox/`
3	Align	One Q&A round — project type, key sources, ownership boundaries. Architecture drafted for your review
4	Schedule	Pick maintenance cadence in plain English: "audit twice a week", "dispatcher at 08:30", "deep daydream off"
5	Running	Daily digest. Local dashboard. `/context bundle <task>` ready. Context compounds from here — the system maintains itself.

Simple, guided, automatic. Claude walks you through each step — you review, you decide. Nothing deleted along the way — originals preserved or archived, never silently overwritten.

What it does

Stops context rot in three ways:

	What it does
Organize	Every doc declares what it owns. CLEAR boundaries prevent the same fact living in 5 places.
Maintain	Scheduled jobs check freshness, surface stale docs, propagate updates across linked content.
Use	`/context search` and `/context bundle <task>` give Claude curated, freshness-aware, source-of-truth-ranked context for the job at hand.

The folder tells the LLM what to trust. Claude reads the path before the content: docs/*.md = canonical, _planned/ = not yet real, _archive/ = history, _inbox/ = unprocessed, _collections/ = verbatim evidence, _wiki/ = explanatory. The hierarchy is the trust signal — fewer tokens spent re-deciding, no chance of treating an archived draft as today's truth.

docs/
├── *.md                  ← canonical data points (one source of truth per topic)
├── document-index.md     ← auto-generated TOC
├── current-state.md      ← what's true right now + what changed recently
├── table-of-context.md   ← high-level cluster map for orientation
├── _inbox/               ← drop zone — Claude triages on demand
├── _planned/             ← polished ideas, not yet real
├── _archive/             ← superseded — kept, not deleted
├── _collections/         ← verbatim evidence (invoices, contracts, transcripts)
│   └── <type>/_manifest.md   (each collection has a manifest)
├── _wiki/                ← derivative knowledge (how-tos, source summaries)
│   ├── pages/            ← user-authored explainers
│   ├── source-summary/   ← external captures with citation banners
│   ├── raw/              ← original artifacts (PDFs, HTML)
│   └── queue.md          ← URLs awaiting fetch
└── _bcos-framework/      ← framework docs — never user content
    ├── architecture/     ← 12 zone + system architecture docs
    ├── methodology/      ← 5 CLEAR + ownership + standards docs
    ├── patterns/         ← 8 project-type starting templates
    └── templates/        ← 15 frontmatter / schema / starter files

The wiki zone

Inspired by Karpathy's LLM Wiki approach — give the LLM a wiki of your knowledge, the way you'd brief a smart new hire.

Save anything to the wiki, from anywhere:

/wiki run https://stripe.com/docs/api          ← fetch a URL, summarize, cite
/wiki create from /Users/me/Downloads/spec.pdf ← extract a PDF, link to the binary
/wiki promote docs/_inbox/meeting-notes.md     ← turn an inbox capture into a wiki page

The wiki holds how-tos, runbooks, glossaries, source summaries, post-mortems, decision logs — derivative knowledge that explains and supports your canonical data points. Every page cites what it builds on. Stale pages surface automatically when their sources advance.

Wiki content is searchable alongside the rest of your context:

/context search "stripe billing"       ← cross-zone search, ranked by source-of-truth
/context bundle market-report:write    ← curated bundle for a specific task

The dashboard

A local cockpit at http://127.0.0.1:8091 that shows your system at a glance — no third-party services, no telemetry, no logins.

Image: BCOS dashboard cockpit — knowledge system status, things waiting on you, maintenance routine

python .claude/scripts/bcos-dashboard/run.py

Surface	What you see
Cockpit	One-line status (healthy / N waiting / heads-up). Per-job dot strip for every maintenance job (core BCOS + wiki zone).
Update	"Run full sync" button → `update.py` + CLAUDE.md drift review + auto-commit + push, with a live log drawer.
Schedules (`/settings/schedules`)	Per-job preset buttons (`daily`, `mon_wed_fri`, `weekly_mon`, `off`) + the global auto-commit toggle + auto-fix whitelist. Saves directly to `schedule-config.json`.
Run history (`/settings/runs`)	Last 20 dispatcher runs with filter chips (job / verdict / trigger).
File health (`/settings/files`)	Frontmatter / cross-reference / staleness findings with one-click fixes.

The dashboard is read-mostly: data flows from the dispatcher's diary, the schedule config, and the canonical context index. Writes (mark-done, schedule preset, auto-commit toggle, file fix) go through dedicated POST endpoints.

Requirements

Claude Code (desktop app, CLI, or web)
Python 3.8+ on PATH
bash (default macOS bash 3.2 is fine)

Updating

python .claude/scripts/update.py

Refreshes framework files only — skills, hooks, scripts, templates. Your docs/ and .private/ are never touched.

Maintenance

One scheduled task per repo. The dispatcher reads schedule-config.json, runs the jobs due today, writes one consolidated digest at docs/_inbox/daily-digest.md, and surfaces findings as one-click cards in the dashboard.

Image: BCOS maintenance schedule dispatcher — daily, weekly, monthly jobs flowing into the daily digest and dashboard cards

You can tune any of it in plain English:

"Run the audit twice a week." · "Move dispatcher to 08:30." · "Turn off deep daydream."

The system learns from your clicks. After three consistent uses of the same fix, the dashboard pre-selects it as a "✨ suggested" action — you accept or reject. Reversals trigger an automatic safety brake.

CLEAR — the methodology

Five principles, one rule each:

C — Contextual Ownership · one document owns each topic. No duplicates.
L — Linking · reference sources; don't copy them.
E — Elimination · consolidate; don't keep two versions.
A — Alignment · context that doesn't support decisions doesn't belong.
R — Refinement · structured maintenance, not ad-hoc fixes.

The whole framework is the mechanical implementation of these five rules.

What ships

Everything below installs as one repo drop-in. No third-party services, no logins, no telemetry — all state lives in your .claude/ and docs/ folders.

Layer	Count	Examples
Skills (chat-invokable via slash command)	17	`clear-planner` · `bcos-wiki` · `context-routing` · `context-audit` · `context-ingest` · `daydream` · `learning` · `schedule-tune` · `lessons-consolidate` · `ecosystem-manager`
Sub-agents (delegated bounded tasks)	3	`wiki-fetch` · `explore` · plus one optional
Scripts (mechanical Python)	46	`lifecycle_sweep` · `run_wiki_` (5) · `cmd_wiki_` (7) · `auto_fix_audit` · `bcos_inventory` · `digest_sidecar` · `record_resolution`
Hooks (auto-validation on edit)	4	post-edit frontmatter check · post-commit context check · pre-compact save · auto-save session
Maintenance jobs (scheduled, one dispatcher)	12	`index-health` · `audit-inbox` · `daydream-lessons` · `daydream-deep` · `architecture-review` · `auto-fix-audit` · `lifecycle-sweep` · `wiki-stale-propagation` · `wiki-source-refresh` · `wiki-graveyard` · `wiki-coverage-audit` · `wiki-canonical-drift`
Auto-fix IDs (silent dispatcher fixes)	12	`missing-last-updated` · `eof-newline` · `frontmatter-field-order` · `wiki-mark-queue-ingested` · `wiki-archive-expired-post-mortem`
Architecture docs	12	`wiki-zone.md` · `lifecycle-routing.md` · `typed-events.md` · `context-zones.md` · `content-routing.md` · `wiki-headless-scripts.md`
Methodology docs	5	CLEAR principles · context architecture · decision framework · document standards · ownership specification
Project patterns (starting templates by project type)	8	client-project · gtm · marketing · operational · product-development · product-service · internal-tool-app · internal-tool-automation
Templates (frontmatter, schema, starters)	15	data-point · cluster · current-state · table-of-context · wiki-schema · wiki-config · session-diary-starter

Every script and skill is documented at the source. Every job has a reference doc under .claude/skills/schedule-dispatcher/references/job-*.md. Every typed-event finding maps to a renderer in the dashboard cockpit. Adding a new check means dropping a cmd_*.py or run_*.py, registering it in one map, and the dashboard picks it up.

Contributing

Contributions welcome — open an issue or submit a PR targeting the dev branch. CI runs frontmatter, reference, and ecosystem-integration checks on every PR.

The fastest contribution path is lessons learned. As you use the system it captures insights; if you find one universal, contribute it back.

Origin

Built by Guntis Coders after two years of operating context systems in production. The CLEAR methodology emerged from a practical problem: AI context degrades silently, and no amount of one-time documentation prevents it. The solution required a system — ownership boundaries, automated maintenance, self-learning, and structured reflection.

Background: What is the Table of Contexts and why does it matter.

License

MIT

docs/_bcos-framework/architecture/context-routing.md

Context Routing Architecture

How CLEAR Context OS turns "I'm about to do task T" into a curated, freshness-flagged, source-of-truth-aware bundle of context. Routing is a derived layer on top of context-zones.md (zones), metadata-system.md (per-doc metadata), and the typed-edge graph in context-index.json (builds-on, references, depends-on, provides).

For the user-facing surface see context-routing/SKILL.md and bundle.md. For zone-scoped sugar over the same backend see bcos-wiki/bundle.md.

Net summary in one line: a task profile names a job-to-be-done (e.g., market-report:write) and declares the shape of context that job needs — required zones, content families, source-of-truth ranking, freshness thresholds, traversal hints, coverage assertions. The bundle resolver walks the typed-edge graph at runtime, applies the rules deterministically, and returns a structured envelope. Mechanical-first (D-10 strict): LLM is opt-in via --resolve-conflicts / --verify-coverage flags only.

Why Routing Exists

/context search "stripe billing" answers "what do we know about X?" — a great keyword retrieval surface (P2). But agents writing market reports, planning customer engagements, or logging architectural decisions don't have a query string in mind. They have a task, and the task implicitly demands:

which zones to pull from (active for canonical truth, planned for proposals, collections for evidence, framework for spec, …),
which content families must be present (competitor data, pricing data, market data),
who wins when canonical and derived disagree,
how fresh each zone must be for this task,
how deep to walk relationships (builds-on two hops, references one hop),
what coverage counts as "enough" (≥1 hit per family).

A keyword query can't carry that. A profile can.

The Profile

A task profile is a declarative document with the following shape (per the catalog at _context.task-profiles.yml.tmpl):

Field	Meaning	Example
`id`	Stable profile identifier	`market-report:write`
`description`	One-line human description	`Writing a market report. …`
`required-zones`	Inline list of `<zone-id>=<true	false>`;` true`= required,`false` = optional
`content-families`	Block list of named families with patterns + min-counts	`name=competitor-data, pattern=cluster=Competitive, required=true, min-count=1`
`source-of-truth-ranking`	Ordered list of zone IDs; first wins on conflict resolution	`[active, wiki, planned]`
`freshness-thresholds`	Per-zone `<zone-id>=<days	never>`; hits past threshold get verdict` past-threshold`
`traversal-hints`	Edge kinds and depth caps to walk from seed hits	`[from-edge=builds-on, depth-cap=2]`
`coverage-assertions`	Per-family `<name>=<min-count>`; failures land in `missing-perspectives`	`[competitor-data=1]`

Pattern grammar (used by content-families and applied to context-index.json doc records):

Pattern	Matches when
`cluster=<value>`	`doc.cluster == <value>`
`tag=<value>`	`<value>` ∈ `doc.tags`
`page-type=<value>`	`doc.page_type == <value>`
`type=<value>`	`doc.type == <value>`

Profiles describe the shape of context needed (patterns, thresholds, ranking), never the instances (specific slugs). The bundle resolver walks the typed-edge graph at runtime; profiles are stable across reorgs.

The Catalog

Ten profiles ship with the framework template:

Profile	When	Notable rule
`market-report:write`	Writing a market report	Canonical wins on conflict; pulls competitor + market + pricing
`engagement:plan`	Planning customer engagement	Live data preferred; wiki supports but never beats canonical
`architecture:design`	Designing or revising architecture	Framework spec + active data points anchor; wiki explainers support
`competitor:audit`	Running a competitor audit	Wiki source-summary + collections heavy; active anchors
`customer:onboard`	Customer onboarding	People + engagement context; live-data-leaning
`decision:log`	Logging an architectural / business decision	Frozen-at-time evidence wins; live data is supportive
`incident:postmortem`	Incident postmortem	Frozen-at-time evidence + decision context
`skill:author`	Authoring or modifying a skill	Framework spec authoritative; ecosystem state load-bearing
`plan:revise`	Revising a planned proposal	Other plans + active data points
`value-lookup:answer`	Answering measurement-shape questions (how much / when / who / top-N / biggest)	Evidence wins over taxonomy: collection-sidecar ranked highest, then manifest, then artifact, then active. For when the question shape is measurement, not definition.

The catalog is loaded by load_task_profiles.py; validated by validate_task_profiles.py (wired into FIXED END / doc-lint). Per-repo override at docs/.context.task-profiles.yml.

Catalog merge semantics

When both the framework template and the per-repo override exist, the loader merges them — the override no longer replaces the template wholesale:

Profile ID is…	Result
In template only	Kept as-is from the framework.
In override only	Added to the catalog after the framework profiles.
In both	Per-repo override wins — its entry replaces the framework's at the same position.

Migration note for existing plugins. If your docs/.context.task-profiles.yml exists only because you needed to add a single profile alongside the framework's nine, you can now drop everything except the profile(s) you actually customise. Re-syncing on framework updates is no longer necessary — the merge handles it. (Deep-merging within a profile is intentionally not supported: redeclare the whole profile by ID to change it.)

The Resolver

context_bundle.py is the resolver. Single entry point: resolve_bundle(profile_id, ...) returns the envelope.

Resolution steps (mechanical, deterministic)

Load profile + corpus. Profile via load_task_profiles; corpus via context-index.json (built by context_index.py).
Build candidate pool. Docs whose zone appears in required-zones.
Walk typed edges. From every seed in the pool, BFS over edges of the kinds declared in traversal-hints up to depth-cap. Each hop records (from, edge, to, depth) in traversal-hops.
Group by zone. Every pool hit lands in by-zone[<zone-id>].
Group by family. Apply each family's pattern; matching hits land in by-family[<family>].
Compute freshness verdict. Per-hit fresh / stale / past-threshold / unknown against the profile's per-zone thresholds. never-zones (e.g., archive) always return fresh.
Detect conflicts. Hits in the same family that share ≥1 EXCLUSIVELY_OWNS key across different zones are CLEAR-violation candidates. Resolved by source-of-truth-ranking: highest-ranked zone wins. Ties at the top zone-rank → resolved-by: unresolved.
Compute coverage gaps. Families failing coverage-assertions[<family>] min-count → missing-perspectives.
Surface unsatisfied zones. required: true zones with no docs → unsatisfied-zone-requirements.

Same fixture index → byte-identical envelope (modulo generated-at). Latency target: <1 second on a 200-doc corpus.

Output envelope

{
  "profile-id": "market-report:write",
  "generated-at": "2026-05-04T19:00:00Z",
  "by-zone": { "active": [...], "wiki": [...], "planned": [...] },
  "by-family": { "competitor-data": [...], "market-data": [...], "pricing-data": [...] },
  "freshness": [
    { "path": "docs/competitors.md", "verdict": "past-threshold", "days-since": 80, "threshold": 30 }
  ],
  "source-of-truth-conflicts": [
    {
      "family": "pricing-data",
      "shared-owns": "tier-names",
      "candidates": [
        { "path": "docs/pricing.md", "zone": "active", ... },
        { "path": "docs/_planned/pricing-redesign.md", "zone": "planned", ... }
      ],
      "resolution": "docs/pricing.md",
      "resolved-by": "rank",
      "reason": "highest-ranked zone: active"
    }
  ],
  "missing-perspectives": [],
  "traversal-hops": [
    { "from": "docs/_wiki/pages/stripe-integration.md", "edge": "builds-on", "to": "docs/pricing.md", "depth": 1 }
  ],
  "unsatisfied-zone-requirements": [],
  "escalations": []
}

A hit (per-doc summary inside by-zone / by-family / candidates) carries: path, name, zone, cluster, page-type, type, tags, exclusively_owns, last-updated, last-reviewed, age_days. Other context-index fields are intentionally trimmed to keep the envelope small.

Source-of-Truth Ranking — How Conflict Resolution Works

The CLEAR principle: every fact has exactly one canonical home. When two docs declare the same EXCLUSIVELY_OWNS key, that's either a real boundary violation or a forking proposal in _planned/. Both cases must surface to the human; the routing layer's job is to flag them, not silence them.

The resolver:

Indexes hits by EXCLUSIVELY_OWNS keys per family.
Any owns-key with ≥2 hits across different zones is a candidate conflict.
Resolution uses the profile's source-of-truth-ranking: the candidate whose zone appears earliest wins. Ties at the top zone-rank → unresolved.
Same-zone duplicates are NOT conflicts (those should already trip the duplication-vs-data-point lint before the resolver runs).

Per-profile ranking is intentional — decision:log ranks collection-artifact (frozen evidence) above active; engagement:plan does the opposite. Ranking is a property of the task, not a global constant.

Freshness Verdicts

Per-hit verdict against the profile's per-zone threshold:

Verdict	When
`fresh`	Threshold is `never`, or `age_days <= threshold/2`
`stale`	`threshold/2 < age_days <= threshold`
`past-threshold`	`age_days > threshold`
`unknown`	`age_days` is null (no `last-updated` field)

The freshness-field per zone is determined by the zone registry: active uses last-updated, wiki pages use last-reviewed, wiki source-summary uses last-fetched (via the freshness-by-page-type override).

D-10 Strict — No Auto-Trigger

The default mechanical run never invokes the LLM. Two opt-in flags exist; both currently require --dry-run until the LLM client wiring lands:

Flag	What it would do (when wired)	Today
`--resolve-conflicts`	LLM picks winners among `resolved-by: unresolved` candidates	Records `resolve-conflicts-dry-run` in `escalations`; non-dry-run raises `LLMEscalationNotImplementedError`
`--verify-coverage`	LLM verifies prose-level coverage of declared perspectives	Records `verify-coverage-dry-run`; non-dry-run raises

The principle is the same as /context search --semantic (P2): mechanical default; explicit flag is the only escalation path; non-dry-run fails loudly rather than silently returning the mechanical envelope while pretending to have escalated.

Determinism and Cost

Determinism. Same fixture index → byte-identical envelope (modulo generated-at). Bundle output is suitable for diffing across runs to detect drift.
Cost. O(N × F) for family grouping (N = pool size, F = profile families). Edge walking is BFS bounded by depth-cap; per-seed work is O(E) over outgoing edges. Conflict detection is O(F × O²) where O is owns-keys per family — for typical profiles this is well under a millisecond.
No LLM cost unless the user opts in.

How Consumers Use the Bundle

Typical agent flow:

The user kicks off a task: "Write me a market report."
The agent runs /context bundle market-report:write --json.
The agent reads the envelope and decides what to actually load:
- Walks by-zone["active"] first (canonical truth).
- Pulls by-family["competitor-data"] to ground the competitor section.
- Notes freshness — anything past-threshold gets a "this hit is stale" annotation in the output.
- Surfaces source-of-truth-conflicts to the user before drafting: "pricing-data has a conflict between docs/pricing.md (active) and docs/_planned/pricing-redesign.md (planned). Treating active as authoritative; flagging the planned proposal."
- Treats missing-perspectives as gaps to call out: "I don't have any market-data hits. The report will reflect this."
- Uses traversal-hops to follow builds-on chains where appropriate.

The bundle is reference data, not a directive. The agent decides what to actually pull into context based on the envelope — which is the whole point of keeping the envelope small.

How Routing Composes With the Rest of the Architecture

context-zones.md — zones, freshness models, source-of-truth roles. Routing reads the registry to know what zones exist and how to address them.
metadata-system.md — per-doc fields. Routing uses cluster, tags, page_type, type, exclusively_owns, last-updated, last-reviewed.
wiki-zone.md — wiki-zone specifics including freshness-by-page-type. Routing reads freshness-thresholds accordingly.
collections-zone.md — collections (evidence) zone. decision:log and incident:postmortem profiles rank collection-artifact highly.
content-routing.md — write-side routing (where new content lands). Distinct from the read-side routing this doc covers.
system-design.md — overall ecosystem map; references both context-zones and context-routing.

Tests

test_context_capability.py — 20 mechanical assertions:

Loader returns normalized profile entries.
Validator rejects unknown zones, missing IDs, malformed entries.
Envelope contains every required key.
by-zone groups hits correctly.
by-family applies pattern matching correctly.
Conflict detection on overlapping EXCLUSIVELY_OWNS across zones.
Conflict resolution picks the highest-ranked zone.
Freshness verdicts use per-zone thresholds.
Coverage gaps surface only when min-count unmet.
Traversal walks builds-on edges to depth-cap.
Determinism: two runs produce identical envelopes (modulo generated-at).
Empty corpus surfaces unsatisfied-zone-requirements.
D-10: default doesn't escalate; opt-in flags are the only LLM trigger.
Non-dry-run opt-in raises LLMEscalationNotImplementedError.
CLI smoke against the fixture profile + corpus.

Summary

A profile names a task and declares the shape of context it needs.
The resolver walks the typed-edge graph deterministically and returns a structured envelope.
Conflicts surface mechanically; resolution uses per-profile ranking; ties stay unresolved for the human.
Freshness is per-zone, threshold-driven, with never as a first-class option.
LLM is opt-in only via explicit flags. The default is honest mechanical Python.
Profiles describe shapes, not instances. Reorgs don't break them.

The bundle is the layer that makes wiki + plans + context docs + collections + lessons + skills + registries function as a coordinated context base, not a pile of zones. Without P5, retrieval is keyword-search-and-pray; with P5, an agent declares its task and gets a curated, ranked, freshness-flagged, gap-aware bundle.

278 files · 9 not served here

Advantages

Prevents context rot by surfacing stale information on a configurable schedule. evidence: README.md
Enforces ownership boundaries to ensure one source of truth per topic. evidence: README.md
Reduces token usage by providing a structured index that avoids redundant scanning. evidence: CLAUDE.md
Enables task-aware context retrieval through curated bundles. evidence: README.md

Limitations

Requires adherence to a specific folder structure and frontmatter schema to function correctly. evidence: README.md
Maintenance jobs and dispatcher functionality rely on Python and bash being available on the host system. evidence: README.md

Security

Risk level: Medium risk

This brain runs

Agent instructions that run or fetch something (170)

.claude/agents/wiki-fetch/AGENT.md: A 100k-token URL fetched into the main thread by `/wiki run` saturates the context window before a single page lands.
.claude/agents/wiki-fetch/AGENT.md: | Calling skill | Subcommand | Trigger | |---|---|---| | `bcos-wiki` | `run` Pass 1 | Each URL in the queue → one `wiki-fetch` invocation | | `bcos-wiki` | `refresh` full-tier | Source-summary past `stale_threshold_days` and HEAD signals changed | | `bcos-wiki` | `promote` (Path B) | S …[+132 chars] · uses the network
.claude/agents/wiki-fetch/AGENT.md: The cumulative budget guard ([`_wiki_budget.decide_dispatch_strategy`](../../scripts/_wiki_budget.py)) decides whether N sub-agents run in parallel or serial — `parallel` when projected cumulative tokens stay ≤ 80% of the main context limit;
.claude/agents/wiki-fetch/AGENT.md: queue an explicit retry tag for next run | | `error.kind: fetch-error` | 4xx/5xx from upstream | Surface to user;
.claude/agents/wiki-fetch/AGENT.md: `python .claude/scripts/test_wiki_fetch.py` — asserts this AGENT.md declares the documented input kinds and output fields, that the 4000-token cap is stated, that calling-skill docs reference both `wiki-fetch` and the Task tool, and that the result-schema validator + cumulative budget g …[+47 chars]
.claude/skills/bcos-wiki/SKILL.md: **`/wiki`** with subcommands: `init`, `init --defaults`, `run`, `queue`, `promote`, `create`, `review`, `archive`, `refresh`, `lint`, `remove`, `schema`, `search`
.claude/skills/bcos-wiki/SKILL.md: "Run the wiki refresh"
.claude/skills/bcos-wiki/SKILL.md: | | `run [<url>]` | implemented | P3 | Path A: batch (no arg) processes all `## Pending` queue items;
.claude/skills/bcos-wiki/SKILL.md: single (with URL) auto-queues + ingests | | `queue <url> [tags]` | implemented | P3 | Append URL to `## Pending` of `queue.md` without fetching | | `ingest` | shared protocol | P3 | Internal — called by `run`/`promote`/`create`;
.claude/skills/bcos-wiki/SKILL.md: optionally re-pull `builds-on` data points to verify | | `archive <slug>` | implemented | P3 | Soft-delete to `_wiki/.archive/` with dangling-reference scan | | `refresh <slug>` | implemented | P3 | Path A: re-fetch a `source-summary` page (refresh-must-rediscover applies, D-05 two-tie …[+470 chars]
.claude/skills/bcos-wiki/SKILL.md: `run` → `run.md`
.claude/skills/bcos-wiki/SKILL.md: Run `/wiki init` first."* and exit.
.claude/skills/bcos-wiki/SKILL.md: | Choice | Behavior | |---|---| | Init with defaults + proceed | Run `.claude/scripts/cmd_wiki_init.py` (or equivalent `ensure_wiki_zone.py` call).
.claude/skills/bcos-wiki/SKILL.md: Then resume the original subcommand (`create from`, `promote`, `run`, etc.) as if the Guard had passed.
.claude/skills/bcos-wiki/SKILL.md: | | Full interview + proceed | Read `init.md` and run the 6-question interactive scaffold.
.claude/skills/bcos-wiki/SKILL.md: The dispatched subcommand files (`create.md`, `promote.md`, `run.md`, `refresh.md`, `lint.md`, `archive.md`, `review.md`, `remove.md`, `schema.md`, `search.md`, `bundle.md`, `queue.md`) all reference *"the `.config.yml` Guard defined in `SKILL.md`"* — they inherit this AskUserQuestion p …[+46 chars]
.claude/skills/bcos-wiki/SKILL.md: The human reviews diffs and runs `git commit` / `git push` when ready. · uses the network
.claude/skills/bcos-wiki/SKILL.md: This applies to `run`, `refresh`, and `create` (when source content is large).
.claude/skills/bcos-wiki/SKILL.md: Users who want a binary in collections-as-evidence must run `/collections add` explicitly.
.claude/skills/bcos-wiki/SKILL.md: ├── run.md ← /wiki run [<url>]
.claude/skills/bcos-wiki/SKILL.md: ├── ingest.md ← shared ingest pipeline (called by run/promote/create/refresh)
.claude/skills/clear-planner/SKILL.md: Step 2: Run Discovery
.claude/skills/clear-planner/SKILL.md: { "step": 2, "name": "Run Discovery", "status": "in_progress" }
.claude/skills/clear-planner/SKILL.md: **Run all three discovery scripts:**
.claude/skills/clear-planner/SKILL.md: bash .claude/agents/agent-discovery/find_agents.sh
.claude/skills/clear-planner/SKILL.md: bash .claude/skills/skill-discovery/find_skills.sh
.claude/skills/clear-planner/SKILL.md: python .claude/scripts/find_lessons.py --tags "#{scenarioType}"
.claude/skills/clear-planner/SKILL.md: Run discovery scripts (Step 2 results)
.claude/skills/clear-planner/SKILL.md: | | ------------------- | ------------------------------------ | --------------------------------------------------------- | ------------------- | | Integration Audit | Detect stale references in ecosystem | `python .claude/scripts/analyze_integration.py --staged` + AI review | **YES …[+455 chars]
.claude/skills/clear-planner/SKILL.md: | Scenario | Key Requirement | | ----------------- | ------------------------------------------------------------------ | | **AGENTING** | Check for ecosystem overlaps + run integration audit before commit | | **DOCUMENTAT …[+76 chars]
.claude/skills/context-audit/SKILL.md: Run all five audit categories against the scoped files.
.claude/skills/context-audit/SKILL.md: python .claude/scripts/build_document_index.py
.claude/skills/context-ingest/SKILL.md: | Input shape | Dispatch to | Why | |-------------|-------------|-----| | HTTP/HTTPS URL (article, GitHub, YouTube) | `/wiki run <url>` (the `bcos-wiki` skill, Path A) | Wiki owns external-source ingest, banner citations, schema validation, refresh tier | | Path inside `docs/_inbox/` a …[+508 chars]
.claude/skills/context-ingest/SKILL.md: If the input is mixed (e.g. Slack export *and* a URL), do mine first, then handle the URL via `/wiki run` separately, then return to ingest for whatever's left.
.claude/skills/context-ingest/SKILL.md: Run `python .claude/scripts/build_document_index.py` after material changes
.claude/skills/context-ingest/SKILL.md: **→ Rebuild Document Index** — Always: `python .claude/scripts/build_document_index.py`
.claude/skills/context-ingest/SKILL.md: **→ context-audit** — If a new data point was created: "Want me to run a quick CLEAR audit on the new doc?"
.claude/skills/context-mine/SKILL.md: Recommend: "Run context-ingest to classify and route these items"
.claude/skills/context-onboarding/SKILL.md: | What you see | What you do | |---|---| | User shares a URL, pitch deck, or describes their business | Gather from those sources, ask 2-3 follow-ups if needed | | Repo already has docs/, README, markdown files with business content | Scan the repo, map what exists | | User drops files …[+394 chars] · uses the network
.claude/skills/context-onboarding/SKILL.md: Before anything else, run **once** per onboarding session to make sure the BCOS-managed instructions are present in `CLAUDE.md`.
.claude/skills/context-onboarding/SKILL.md: This catches the `git clone`-without-`install.sh` case: if `CLAUDE.md` exists but lacks the `` / `` markers, the framework section is missing or unstructured. · uses the network
.claude/skills/context-onboarding/SKILL.md: If `.claude/bcos-claude-reference.md` doesn't exist yet (fresh clone), point `--source` at the upstream template available in this repo's tree (the same file install.sh would have copied).
.claude/skills/context-onboarding/SKILL.md: **First-run routing principle:** onboarding is allowed to identify all BCOS zones, but it should only create active data points after the user approves the architecture.
.claude/skills/context-onboarding/SKILL.md: | Material | First-run handling | |---|---| | Canonical facts, current decisions, project/task/business reality | Draft active data points in `docs/*.md` | | Raw notes, unsorted snippets, uncertain material | Keep in `docs/_inbox/` until triaged | | Future ideas, plans, or maybes | Par …[+578 chars]
.claude/skills/context-onboarding/SKILL.md: Options: **Activate** (set all to active, run index + wake-up) / **Adjust** (tell me what to fix)
.claude/skills/context-onboarding/SKILL.md: Run `python .claude/scripts/build_document_index.py` to generate the Document Index with DOMAIN one-liners
.claude/skills/context-onboarding/SKILL.md: Run `python .claude/scripts/generate_wakeup_context.py` to generate the Wake-Up Context for instant session orientation
.claude/skills/context-onboarding/SKILL.md: The checklist ships with install at `docs/.onboarding-checklist.md`.
.claude/skills/context-onboarding/SKILL.md: Ensure the `.claude/hook_state/` directory exists (it usually does from install.sh — check and create if missing).
.claude/skills/context-onboarding/SKILL.md: the dispatcher will create it on first run.
.claude/skills/context-onboarding/SKILL.md: **Run the catalog drift guard:**
.claude/skills/context-onboarding/SKILL.md: python .claude/scripts/validate_permissions_catalog.py
.claude/skills/context-onboarding/SKILL.md: python .claude/scripts/update.py
.claude/skills/context-onboarding/SKILL.md: Re-run the validator to confirm clean.
.claude/skills/context-onboarding/SKILL.md: If the answer is "Yes, multi-repo" (or the user confirms after explanation), run:
.claude/skills/context-onboarding/SKILL.md: python .claude/scripts/install_global_permissions.py
.claude/skills/context-onboarding/SKILL.md: Question: "When should the morning dispatcher run?"
.claude/skills/context-onboarding/SKILL.md: Run the schedule-dispatcher skill to execute today's scheduled CLEAR maintenance. It will:
.claude/skills/context-onboarding/SKILL.md: 3. Execute each job in sequence, appending one diary entry per job
.claude/skills/context-onboarding/SKILL.md: The portfolio aggregator picks it up on the next dispatcher run.
.claude/skills/context-onboarding/SKILL.md: The dispatcher will run in the wrong directory and writes meant for this repo's `docs/_inbox/` will be blocked.
.claude/skills/context-onboarding/SKILL.md: If "Recreate", call `mcp__scheduled-tasks__delete_event` (or the equivalent delete tool) on the bad task, then re-run Step 6d.
.claude/skills/context-onboarding/SKILL.md: That's intentional — fresh setups should over-monitor for the first 1-2 weeks until the auto-tuner suggests reductions after green-run streaks.
.claude/skills/context-onboarding/SKILL.md: Let the user invoke the first run on their terms.
.claude/skills/context-onboarding/SKILL.md: A fresh install has nothing to migrate and should never see the word.
.claude/skills/core-discipline/SKILL.md: **Significant change** (restructure a cluster, create new skills) → use clear-planner, run the full workflow
.claude/skills/core-discipline/SKILL.md: | Task Type | Action | | ------------------------- | ---------------------- | | Questions, commands | Respond directly | | Single-file fixes | Just do it | | Small changes (1-3 files) | Think briefly, execute |
.claude/skills/core-discipline/SKILL.md: A scheduled dispatcher run finishes clean green with zero findings (one-line summary keeps the dashboard "Ready" status meaningful)
.claude/skills/core-discipline/SKILL.md: **Special case — wrapping scripts with built-in prompts** (like `update.py`'s `[y/N]` prompt): invoke the script with `--dry-run` first, parse the output, surface the summary via `AskUserQuestion`, then re-invoke with `--yes` based on the user's choice.
.claude/skills/daydream/SKILL.md: If the file doesn't exist, this is the first run — use 2 weeks ago as the baseline.
.claude/skills/daydream/SKILL.md: # Read last run timestamp (or default to 2 weeks ago)
.claude/skills/daydream/SKILL.md: cat .claude/quality/last-daydream.txt 2>/dev/null || echo "no previous run"
.claude/skills/daydream/SKILL.md: git log --since="{last_daydream_date}" --name-only --pretty=format:"%h %s (%ar)" -- docs/ ':!docs/_bcos-framework/methodology/' ':!docs/_bcos-framework/guides/' ':!docs/_bcos-framework/templates/'
.claude/skills/daydream/SKILL.md: Report "No document changes since last run" and skip to Phase 3 (Imagine) for forward-looking reflection only.
.claude/skills/daydream/SKILL.md: git diff {last_daydream_commit}..HEAD -- docs/path/to/changed-file.md
.claude/skills/daydream/SKILL.md: python .claude/scripts/build_document_index.py --dry-run
.claude/skills/daydream/SKILL.md: python .claude/scripts/generate_wakeup_context.py
.claude/skills/daydream/SKILL.md: **Step 5b: Run maintenance scripts**
.claude/skills/daydream/SKILL.md: Daydream is the natural time to run periodic maintenance:
.claude/skills/daydream/SKILL.md: python .claude/scripts/prune_sessions.py
.claude/skills/daydream/SKILL.md: python .claude/scripts/prune_diary.py
.claude/skills/daydream/SKILL.md: python .claude/scripts/generate_wakeup_context.py
.claude/skills/daydream/SKILL.md: python .claude/scripts/analyze_crossrefs.py
.claude/skills/daydream/SKILL.md: **Step 6: Record this run**
.claude/skills/daydream/SKILL.md: python .claude/scripts/record_daydream.py
.claude/skills/daydream/SKILL.md: **Document Index** — Run `python .claude/scripts/build_document_index.py` if files changed
.claude/skills/doc-lint/SKILL.md: If `.md`: Run markdown quality checks and cross-reference validation
.claude/skills/doc-lint/SKILL.md: If `.json`: Run JSON structure validation
.claude/skills/doc-lint/SKILL.md: Px_001: Run markdown quality check -> doc-lint (markdown category)
.claude/skills/doc-lint/SKILL.md: Px_002: Run example code validation -> doc-lint (cross_reference category)
.claude/skills/doc-lint/SKILL.md: Px_001: Run JSON structure validation -> doc-lint (json_structure category)
.claude/skills/doc-lint/SKILL.md: Px_002: Run markdown quality check -> doc-lint (markdown category)
.claude/skills/ecosystem-manager/SKILL.md: Run discovery to see what exists:
.claude/skills/ecosystem-manager/SKILL.md: bash .claude/agents/agent-discovery/find_agents.sh
.claude/skills/ecosystem-manager/SKILL.md: bash .claude/skills/skill-discovery/find_skills.sh
.claude/skills/ecosystem-manager/SKILL.md: Run it before committing.
.claude/skills/ecosystem-manager/SKILL.md: python .claude/scripts/analyze_integration.py --uncommitted
.claude/skills/ecosystem-manager/SKILL.md: Coverage gaps in install.sh, settings.json, state.json, .gitignore
.claude/skills/ecosystem-manager/SKILL.md: Add fix tasks to the plan and execute them before committing.
.claude/skills/ecosystem-manager/SKILL.md: [ ] **install.sh installs new files**
.claude/skills/ecosystem-manager/SKILL.md: Run after major development sessions or as part of monthly architecture review.
.claude/skills/ecosystem-manager/SKILL.md: Run `python .claude/scripts/wiki_schema.py list` to inspect page-types, lint-checks, and auto-fixes from `_wiki/.schema.yml` or the framework fallback.
.claude/skills/ecosystem-manager/SKILL.md: Run all 4 CI checks locally:
.claude/skills/ecosystem-manager/SKILL.md: python .github/scripts/validate_frontmatter.py # YAML metadata on all managed markdown
.claude/skills/ecosystem-manager/SKILL.md: python .github/scripts/validate_references.py # All registry paths resolve + state.json accuracy
.claude/skills/ecosystem-manager/SKILL.md: python .claude/scripts/analyze_integration.py --ci # Ecosystem wiring (install.sh, settings.json coverage)
.claude/skills/learning/SKILL.md: **Mechanism:** `python .claude/scripts/promote_resolutions.py` (the module's `__main__` calls `regenerate()`).
.claude/skills/learning/SKILL.md: Run the matching subcommand directly.
.claude/skills/lessons-consolidate/SKILL.md: python .claude/scripts/consolidate_lessons.py
.claude/skills/schedule-dispatcher/SKILL.md: Run `python .claude/scripts/update.py` first."
.claude/skills/schedule-dispatcher/SKILL.md: Report to the user: "schedule-config.json is malformed — fix before the dispatcher can run.
.claude/skills/schedule-dispatcher/SKILL.md: For each entry in `jobs`, decide if it should run today.
.claude/skills/schedule-dispatcher/SKILL.md: | `schedule` value | Runs when | |--------------------|----------------------------------------------------| | `"daily"` | Every day | | `"mon"`..`"sun"` | Only on that weekday …[+549 chars]
.claude/skills/schedule-dispatcher/SKILL.md: **For on-demand dispatch** (user said "run the audit-inbox job"): skip this step, run only the named job.
.claude/skills/schedule-dispatcher/SKILL.md: Build an ordered list of jobs to run.
.claude/skills/schedule-dispatcher/SKILL.md: python .claude/scripts/context_index.py --write
.claude/skills/schedule-dispatcher/SKILL.md: The cache TTL is 10 min (longer than a full dispatcher cycle), so all jobs in this run hit the same in-memory snapshot.
.claude/skills/schedule-dispatcher/SKILL.md: Last run timestamp per job?
.claude/skills/schedule-dispatcher/SKILL.md: Use the helper script — it creates `.claude/hook_state/` on first run and matches the allowlisted command prefix so it never prompts:
.claude/skills/schedule-dispatcher/SKILL.md: python .claude/scripts/append_diary.py '{"ts":"2026-04-15T09:04:12","job":"index-health","verdict":"green","findings_count":0,"auto_fixed":[],"actions_needed":[],"duration_s":4}'
.claude/skills/schedule-dispatcher/SKILL.md: **Fallback:** when `.bcos-umbrella.json` is absent (standalone BCOS install with no umbrella), preflight reduces to the `../` + absolute-path scans — no sibling-name list to check against.
.claude/skills/schedule-dispatcher/SKILL.md: After Step 4 completes (last job returned), verify that **every job listed in Step 2's run-list produced exactly one diary entry in this dispatcher tick**.
.claude/skills/schedule-dispatcher/SKILL.md: Read back the diary entries appended during this tick (use the dispatcher's in-memory record, or the tail of `.claude/hook_state/schedule-diary.jsonl` filtered to the current run's `ts` prefix if you're running on-demand).
.claude/skills/schedule-dispatcher/SKILL.md: **Added in schema 1.1.0.** Before applying auto-fixes (Step 5) or rendering the digest (Step 7), enrich each finding emitted in Step 4 with three computed fields and produce the run's headline string.
.claude/skills/schedule-dispatcher/SKILL.md: **Match against prior tick.** Scan diary entries from the **immediately-prior dispatcher tick** (the most-recent run before today, identified by the latest `ts` < today's run window).
.claude/skills/schedule-dispatcher/SKILL.md: | Run shape | Headline template | |---|---| | Green clean (0 findings, 0 framework) | "{Nth consecutive green run.}{recovery callout if any job verdict flipped error→green this tick.}" | | Green with frequency suggestions only | "All {N} jobs green.
.claude/skills/schedule-dispatcher/SKILL.md: And the run has a `headline` string ready for Step 7.
.claude/skills/schedule-dispatcher/SKILL.md: For each job run today, check recent diary:
.claude/skills/schedule-dispatcher/SKILL.md: **Last run was `error`** → suggest investigating, not frequency change
.claude/skills/schedule-dispatcher/SKILL.md: Both are overwritten each run.
.claude/skills/schedule-dispatcher/SKILL.md: Conditional sections render ONLY when their counts are non-zero — a green clean run is ~5 lines, an amber run with 5 findings is ~40 lines.
.claude/skills/schedule-dispatcher/SKILL.md: | 1 | `{emitted_by}` | `{primary_attr}` | {one-line meaning} | {`🔁 Nth run` if consecutive_runs ≥ 3, else blank} |
.claude/skills/schedule-dispatcher/SKILL.md: Run at {ISO timestamp} · `.claude/hook_state/schedule-diary.jsonl`
.claude/skills/schedule-dispatcher/SKILL.md: verdict emoji + finding count + inline status markers in Note column | | `💡 Suggestions` | `≥1` `frequency-suggestion` finding | Bullet list | | Footer | Always | Run timestamp, diary path, auto-commit status |
.claude/skills/schedule-dispatcher/SKILL.md: | Scenario | Verdict bar example | Top line | Action hint | |---|---|---|---| | Green clean | `**🟢 BCOS green** · 1 ran · 0 findings · 0 fixed · 4s` | `Headline string from Step 4c, e.g. "10th consecutive green run."` | `→ Nothing to act on.
.claude/skills/schedule-dispatcher/SKILL.md: Run `git rev-parse --abbrev-ref HEAD` to get the current branch.
.claude/skills/schedule-dispatcher/SKILL.md: This is the per-tick safety property: a stray hand-edit to a "generated" file on a day its owning job did not run still blocks auto-commit.
.claude/skills/schedule-dispatcher/SKILL.md: Globs are resolved against `git status --porcelain` paths (same surface the clean-tree rule already inspects).
.claude/skills/schedule-dispatcher/SKILL.md: **Emit the `job-missing-outputs-declaration` framework-finding only when** the job's run would have changed at least one path *outside* `GLOBAL_ALLOWED` (i.e. a path that *could* have been auto-committable had it been declared).
.claude/skills/schedule-dispatcher/SKILL.md: Run `git status --porcelain`.
.claude/skills/schedule-dispatcher/SKILL.md: `git commit -m "bcos: daily maintenance {YYYY-MM-DD}"` — no push, ever.
.claude/skills/schedule-dispatcher/SKILL.md: A BCOS install with no umbrella still writes the log — it just sits locally until somebody (or the dashboard) reads it.
.claude/skills/schedule-dispatcher/SKILL.md: The last thing the dispatcher does depends on whether the run produced anything that needs user judgement.
.claude/skills/schedule-dispatcher/SKILL.md: Otherwise — clean green run, zero findings, zero suggestions — output a short summary and stop.
.claude/skills/schedule-dispatcher/SKILL.md: **Why this rule:** if every run always asked a question, every scheduled run would show as "Awaiting input" in the dashboard even on clean days.
.claude/skills/schedule-dispatcher/SKILL.md: Or even shorter when the user is watching an on-demand run:
.claude/skills/schedule-dispatcher/SKILL.md: ("Do you want architecture review to run weekly?" is too abstract — `schedule-tune` handles that)
.claude/skills/schedule-dispatcher/SKILL.md: **Retry the failed job(s)** → re-run only the errored jobs (on-demand mode), then ask again with the new result
.claude/skills/schedule-dispatcher/SKILL.md: If the user says "run the {job-name} job now" or "run today's maintenance now":
.claude/skills/schedule-dispatcher/SKILL.md: Skip the schedule predicate — run only the named job(s)
.claude/skills/schedule-dispatcher/SKILL.md: ## On-demand run — {HH:MM} ({job-name})
.claude/skills/schedule-dispatcher/SKILL.md: If `daily-digest.md` does not yet exist (no scheduled run happened today), create it with a minimal header before appending the on-demand section.
.claude/skills/schedule-dispatcher/SKILL.md: DO append a diary entry per job with `"trigger":"on-demand"` so the run is tracked in history
.claude/skills/schedule-dispatcher/SKILL.md: Run another job
.claude/skills/schedule-dispatcher/SKILL.md: The morning scheduled run always OVERWRITES `daily-digest.md` fresh — so yesterday's on-demand sections are not preserved in the digest (diary retains them).
.claude/skills/schedule-dispatcher/SKILL.md: Every dispatcher run produces either a digest OR a user-facing error message.
.claude/skills/schedule-dispatcher/SKILL.md: `schedule-tune` — user-facing skill for changing config ("run audit twice a week")
.claude/skills/schedule-dispatcher/SKILL.md: If a run was wrong, add a corrective entry, don't mutate history.
.claude/skills/schedule-tune/SKILL.md: If you haven't set up BCOS scheduling yet, run `context-onboarding` Step 6.
.claude/skills/schedule-tune/SKILL.md: **Utterances:** *"run audit twice a week"*, *"change daydream to every 3 days"*, *"switch architecture review to quarterly"*, *"run index-health on weekdays only"*, *"audit on Tuesdays and Fridays"*
.claude/skills/schedule-tune/SKILL.md: **Utterances:** *"run the dispatcher at 08:30"*, *"move morning maintenance to 07:00"*
.claude/skills/schedule-tune/SKILL.md: Update the actual scheduled task's cron via `mcp__scheduled-tasks__update_scheduled_task` — that's what actually moves the run
.claude/skills/schedule-tune/SKILL.md: **Utterances:** *"what's my maintenance schedule?"*, *"when does audit run?"*, *"is architecture-review enabled?"*
.claude/skills/schedule-tune/SKILL.md: If the user proposes a cron that would run more than once per day, refuse: *"Dispatcher is designed to run once daily.
.claude/skills/schedule-tune/SKILL.md: | User says | Config edit | |----------------------------------------|----------------------------------------------------------------| | "run audit on Tuesday instead of Friday" | `jobs.audit-inbox.schedu …[+1012 chars]
.claude/skills/todo-utilities/SKILL.md: **Sequential Agent Invocation** — Tasks run one after another
CLAUDE.md: **Small changes (1-3 files)** — think briefly, execute.
CLAUDE.md: Auto-generated by `python .claude/scripts/generate_wakeup_context.py`.
CLAUDE.md: When unsure, run `context-ingest` and let it propose the route.
CLAUDE.md: **Post-work handoffs:** After integrating new content, offer to (a) rebuild the document index, (b) update related data points, (c) run an audit if architecture shifted.

Hooks (5)

.claude/settings.json (Stop): bash "$CLAUDE_PROJECT_DIR/.claude/hooks/auto_save_session.sh"
.claude/settings.json (PreCompact): bash "$CLAUDE_PROJECT_DIR/.claude/hooks/precompact_save.sh"
.claude/settings.json (PostToolUse): "$CLAUDE_PROJECT_DIR/.claude/bin/python3" "$CLAUDE_PROJECT_DIR/.claude/hooks/post_edit_frontmatter_check.py"
.claude/settings.json (PostToolUse): "$CLAUDE_PROJECT_DIR/.claude/bin/python3" "$CLAUDE_PROJECT_DIR/.claude/hooks/post_edit_frontmatter_check.py"
.claude/settings.json (PostToolUse): "$CLAUDE_PROJECT_DIR/.claude/bin/python3" "$CLAUDE_PROJECT_DIR/.claude/hooks/post_commit_context_check.py"

Install scripts (1)

install.sh: sh install.sh

Medium risk: runs-instruction-instruction-network (.claude/agents/wiki-fetch/AGENT.md:93)
Medium risk: runs-instruction-instruction-network (.claude/skills/bcos-wiki/SKILL.md:150)
Medium risk: runs-instruction-instruction-network (.claude/skills/context-onboarding/SKILL.md:33)
Medium risk: runs-instruction-instruction-network (.claude/skills/context-onboarding/SKILL.md:50)
Low risk: scripts-present (.claude/agents/agent-discovery/find_agents.sh)
Low risk: scripts-present (.claude/bin/python3)
Low risk: scripts-present (.claude/hooks/auto_save_session.sh)
Low risk: scripts-present (.claude/hooks/post_commit_context_check.py)
Low risk: scripts-present (.claude/hooks/post_edit_frontmatter_check.py)
Low risk: scripts-present (.claude/hooks/precompact_save.sh)
Low risk: scripts-present (.claude/scripts/analyze_crossrefs.py)
Low risk: scripts-present (.claude/scripts/analyze_integration.py)
Low risk: scripts-present (.claude/scripts/append_diary.py)
Low risk: scripts-present (.claude/scripts/audit_node_configs.py)
Low risk: scripts-present (.claude/scripts/auto_fix_audit.py)
Low risk: scripts-present (.claude/scripts/bcos_inventory.py)
Low risk: scripts-present (.claude/scripts/build_document_index.py)
Low risk: scripts-present (.claude/scripts/cmd_wiki_archive.py)
Low risk: scripts-present (.claude/scripts/cmd_wiki_init.py)
Low risk: scripts-present (.claude/scripts/cmd_wiki_lint.py)
Low risk: scripts-present (.claude/scripts/cmd_wiki_queue_add.py)
Low risk: scripts-present (.claude/scripts/cmd_wiki_remove.py)
Low risk: scripts-present (.claude/scripts/cmd_wiki_review.py)
Low risk: scripts-present (.claude/scripts/cmd_wiki_search.py)
Low risk: scripts-present (.claude/scripts/consolidate_lessons.py)
Low risk: scripts-present (.claude/scripts/context_bundle.py)
Low risk: scripts-present (.claude/scripts/context_index.py)
Low risk: scripts-present (.claude/scripts/context_search.py)
Low risk: scripts-present (.claude/scripts/digest_sidecar.py)
Low risk: scripts-present (.claude/scripts/find_lessons.py)
Low risk: scripts-present (.claude/scripts/generate_wakeup_context.py)
Low risk: scripts-present (.claude/scripts/install_global_permissions.py)
Low risk: scripts-present (.claude/scripts/lifecycle_sweep.py)
Low risk: scripts-present (.claude/scripts/load_task_profiles.py)
Low risk: scripts-present (.claude/scripts/load_zone_registry.py)
Low risk: scripts-present (.claude/scripts/promote_resolutions.py)
Low risk: scripts-present (.claude/scripts/prune_diary.py)
Low risk: scripts-present (.claude/scripts/prune_sessions.py)
Low risk: scripts-present (.claude/scripts/record_daydream.py)
Low risk: scripts-present (.claude/scripts/record_resolution.py)
Low risk: scripts-present (.claude/scripts/refresh_ecosystem_state.py)
Low risk: scripts-present (.claude/scripts/refresh_wiki_index.py)
Low risk: scripts-present (.claude/scripts/run_wiki_canonical_drift.py)
Low risk: scripts-present (.claude/scripts/run_wiki_coverage_audit.py)
Low risk: scripts-present (.claude/scripts/run_wiki_graveyard.py)
Low risk: scripts-present (.claude/scripts/run_wiki_source_refresh.py)
Low risk: scripts-present (.claude/scripts/run_wiki_stale_propagation.py)
Low risk: scripts-present (.claude/scripts/scan_docs_structure.py)
Low risk: scripts-present (.claude/scripts/set_profile.sh)
Low risk: scripts-present (.claude/scripts/update.py)