ask

/bedrock:ask — Adaptive Vault Reader

Plugin Paths

Entity definitions and templates are in the plugin directory, not at the vault root. Use the "Base directory for this skill" provided at invocation to resolve the paths:

• Entity definitions: <base_dir>/../../entities/
• Templates: <base_dir>/../../templates/{type}/_template.md
• Plugin CLAUDE.md: <base_dir>/../../CLAUDE.md (already automatically injected into context)

Where <base_dir> is the path provided in "Base directory for this skill".

---

Vault Resolution

Resolve which vault to query. This skill can be invoked from any directory.

Step 1 — Parse --vault flag: Check if the input arguments include --vault <name>. If found, extract the vault name and remove it from the arguments (the remaining text is the question).

Step 2 — Resolve vault path:

1. If --vault <name> was provided: Read the vault registry at <base_dir>/../../vaults.json. Find the entry matching the name. If not found: error — "Vault <name> is not registered. Run /bedrock:vaults to see available vaults." If found: set VAULT_PATH to the entry's path value.

2. If no --vault flag — CWD detection: Read <base_dir>/../../vaults.json. Check if the current working directory is inside any registered vault path (CWD starts with a registered vault's absolute path). If multiple match, use the longest path (most specific). If found: set VAULT_PATH to the matching vault's path.

3. If CWD detection fails — default vault: From the registry, find the vault with "default": true. If found: set VAULT_PATH to the default vault's path.

4. If no resolution: Error — "No vault resolved. Available vaults:" followed by the registry listing. "Use --vault <name> to specify, or run /bedrock:setup to register a vault."

Step 3 — Validate vault path:

test -d "<VAULT_PATH>" && echo "exists" || echo "missing"

If missing: error — "Vault path <VAULT_PATH> does not exist on disk. Run /bedrock:setup to re-register."

Step 4 — Read vault config:

cat <VAULT_PATH>/.bedrock/config.json 2>/dev/null

Extract language and other relevant fields for use in later phases.

From this point forward, ALL vault file operations use <VAULT_PATH> as the root.

• Entity directories: <VAULT_PATH>/actors/, <VAULT_PATH>/people/, etc.
• Graphify output: <VAULT_PATH>/graphify-out/

---

Overview

This skill receives a natural language question and answers it using an adaptive, vault-first approach. It always reads vault content first, then decides whether to escalate to graphify or /teach based on what's actually needed — not what the question looks like in isolation.

You are an adaptive context orchestrator agent. You only READ — never write, edit, or delete files directly.

Writes happen exclusively through /bedrock:teach delegation (which flows through /bedrock:preserve). If the query reveals outdated or missing information and no remote source is available to ingest, suggest that the user run /bedrock:preserve or /bedrock:teach to update the vault.

---

Phase 0 — Read Configuration

0.1 Load config

Read .bedrock/config.json from the vault root:

if [ -f ".bedrock/config.json" ]; then
    cat .bedrock/config.json
else
    echo "config_not_found"
fi

• If config exists: extract the value of query.max_graphify_calls. Store as max_graphify_calls.
• If config does not exist or field is absent: set max_graphify_calls = 3 (default).
• Valid range: 1–5. If the value is outside this range, clamp to the nearest bound and log a warning.

---

Phase 1 — Analyze the Question

1.1 Classify the question

Read the user's question and identify:

1. Mentioned entities — names of systems, people, teams, topics, projects, or discussions. They may appear as:

   • Exact filename (e.g.: "billing-api", "squad-payments")
   • Human-readable name (e.g.: "Billing API", "Squad Payments")
   • Alias or acronym (e.g.: "BillingAPI", "BRB")
   • Contextual reference (e.g.: "the billing service", "the notifications team")

2. Relevant domain(s) — payments, notifications, orders, integrations, checkout, compliance, internal-tools. Infer from the mentioned entities or the question context.

3. Type of information sought:

   • Status/overview — "what is X?", "what's the status of X?"
   • Architecture/stack — "how does X work?", "what's the stack of X?"
   • People/teams — "who owns X?", "who works with Y?"
   • History/decisions — "what was decided about X?", "what happened with Y?"
   • Relationships — "what depends on X?", "how does Y relate to Z?"
   • Deprecation — "what is being deprecated?", "what's the deprecation plan for X?"

1.2 Assess clarity

If the question is too ambiguous to produce a targeted search (e.g.: "tell me everything", "how does the system work?", "what's going on?"), ask for clarification:

"Your question is broad. Can you specify: which system, team, or topic would you like to know more about?"

If the question mentions something that clearly isn't part of the vault (e.g.: something personal, unrelated technology), inform: "I didn't find anything in the vault about this."

1.3 Phase 1 classification result

At the end, you should have:

• search_terms: list of names, aliases, and keywords to search for
• domains: list of relevant domains (may be empty if not identified)
• info_type: classification of the type of information sought
• explicit_entities: entities mentioned directly by name (if any)

---

Phase 2 — Vault-First Search

This phase always runs for every question, regardless of graph availability. It is the foundation of the adaptive approach — read vault content first, decide later.

2.1 Read entity definitions

Use Read to read the entity definition files from the plugin (see "Plugin Paths" section):

• If the question is about a system → read <base_dir>/../../entities/actor.md
• If the question is about a person → read <base_dir>/../../entities/person.md
• If the question is about a team → read <base_dir>/../../entities/team.md
• If the question is about a topic/deprecation → read <base_dir>/../../entities/topic.md
• If the question is about a meeting/decision → read <base_dir>/../../entities/discussion.md
• If the question is about a project/initiative → read <base_dir>/../../entities/project.md
• If you don't know the type → read all entity definitions from the plugin to classify correctly

2.2 Search entities by name and alias

For each search term identified in Phase 1:

Step 1 — Search by filename:

Glob: <VAULT_PATH>/actors/<term>*.md, <VAULT_PATH>/people/<term>*.md, <VAULT_PATH>/teams/<term>*.md,
      <VAULT_PATH>/topics/*<term>*.md, <VAULT_PATH>/discussions/*<term>*.md, <VAULT_PATH>/projects/<term>*.md,
      <VAULT_PATH>/fleeting/*<term>*.md

Step 2 — Search by alias in frontmatter:

Grep: pattern="aliases:.*<term>" in directories: <VAULT_PATH>/actors/, <VAULT_PATH>/people/, <VAULT_PATH>/teams/,
      <VAULT_PATH>/topics/, <VAULT_PATH>/discussions/, <VAULT_PATH>/projects/
      (case-insensitive)

Step 3 — Search by name in frontmatter:

Grep: pattern="name:.*<term>" or pattern="title:.*<term>"
      in the same directories (case-insensitive)

Step 4 — Search by content (fallback): If steps 1-3 did not return sufficient results:

Grep: pattern="<term>" in entity directories (case-insensitive)

2.3 Filter by domain

If domains were identified in Phase 1, filter results:

Grep: pattern="domain/<domain>" in the found files (tags field of frontmatter)

Keep all results, but prioritize those matching the domain.

2.4 Read found entities

For each entity found (limit: 15 entities):

1. Read the frontmatter first (~first 30 lines) to confirm relevance
2. If relevant: read the full file
3. If not relevant (false positive from Grep): discard

Record for each entity read:

• filename, type, name
• wikilinks found in frontmatter and body
• external URLs found in the content (Confluence, Google Docs, GitHub)
• Explicit date in the filename (if any)

2.5 Follow wikilinks (1 level of depth)

For each extracted wikilink that is relevant to the question:

1. Resolve the file: search for <wikilink-name>.md in entity directories

   Glob: <VAULT_PATH>/actors/<name>.md, <VAULT_PATH>/people/<name>.md, <VAULT_PATH>/teams/<name>.md,
         <VAULT_PATH>/topics/*<name>*.md, <VAULT_PATH>/discussions/*<name>*.md, <VAULT_PATH>/projects/<name>.md
   

2. Read the found file (frontmatter + body)

3. Do NOT follow wikilinks from this second level — stop here to avoid context explosion

Relevance criteria for following a wikilink:

• The question is about relationships ("who owns", "what depends on") → follow all
• The question is about status/overview → follow team, people (focal points)
• The question is about history → follow related discussions, topics
• The question is about architecture → follow dependent actors

Limit: Do not read more than 15 entities total (2.4 + 2.5 combined). If the limit is reached, prioritize entities directly mentioned in the question.

2.6 Phase 2 output

At the end of Phase 2, you have:

• A set of vault entities with their full content
• Wikilinks between them (structural relationships)
• External URLs found in their content (Confluence, GDocs, GitHub)
• A sense of whether the vault content covers the question

---

Phase 3 — Context Assessment + Conditional Escalation

This is the core decision point. After reading vault content in Phase 2, assess whether you have enough context to answer the question.

3.1 Self-Assessment

Evaluate the vault content you read in Phase 2 against the original question. Determine one of three outcomes:

vault_sufficient — You have enough information to compose a good answer. Indicators:

• The question is factual/status/ownership and the vault entities contain a clear answer
• Examples: "who owns X", "what's the status of Y", "what team manages Z", "what is X"
• The entities read in Phase 2 directly address the question
• No significant gaps in the information

needs_graphify — The vault content is partial but the knowledge graph could fill the gaps. Indicators:

• The question involves code-level relationships, cross-domain dependencies, or architectural paths
• The vault entities reference systems whose connections aren't explicit in the markdown
• You feel you're missing structural context that the knowledge graph could provide
• Examples: "how does X connect to Y at the code level", "what are the dependencies of X", "trace the data flow from A to B"

needs_remote_content — The vault entities reference external URLs that appear directly relevant to the question, but the content behind those URLs isn't ingested in the vault. Indicators:

• An entity's sources field or body text contains a URL (Confluence, GDocs, GitHub) that likely holds the answer
• The question asks about something documented externally (e.g., "what's the runbook for X" and the entity links to a Confluence page)
• The vault has a pointer to the answer but not the answer itself

Priority when multiple outcomes apply: needs_remote_content > needs_graphify > vault_sufficient

Rationale: remote content must be internalized first for the vault to be complete. Graphify can run on richer data after ingestion. If both apply, handle remote content first, then re-assess whether graphify is still needed.

After determining the outcome:

• vault_sufficient → skip directly to Phase 4 (recency) then Phase 5 (respond)
• needs_graphify → proceed to Phase 3-G
• needs_remote_content → proceed to Phase 3-T

---

Phase 3-G — Graphify Escalation

Execute only when the self-assessment determines needs_graphify.

3-G.0 Check graph availability

if [ -f "<VAULT_PATH>/graphify-out/graph.json" ] && [ -s "<VAULT_PATH>/graphify-out/graph.json" ]; then
    echo "graph_available"
else
    echo "graph_not_available"
fi

If graph_not_available: Display the following warning and skip to Phase 4 with vault-only content:

Warning

Knowledge graph unavailable The knowledge graph is not available (<VAULT_PATH>/graphify-out/graph.json missing or empty). The answer below is based on vault content only — it may be incomplete for this type of question. Run /graphify build to rebuild the graph from the vault's actor repositories.

3-G.1 Formulate graphify calls

Based on the gap between what you have (Phase 2 content) and what the question needs, formulate 1–N graphify calls. Use the same modes as before:

Gap identified	Graphify mode
Need to understand a single entity's code structure	explain "<entity>"
Need to find how two entities connect	path "<entityA>" "<entityB>"
Need broad relationship or dependency context	query "<question about the gap>"

The LLM decides the calls based on what's missing — not from a pre-planned decomposition. Never exceed max_graphify_calls.

3-G.2 Execute graphify calls sequentially

For each call, invoke /graphify via the Skill tool. Append the structured JSON output instruction:

After completing the traversal, return ONLY a JSON object with this structure (no prose, no markdown fences):
{
  "mode": "query|path|explain",
  "start_nodes": ["node_id1", "node_id2"],
  "nodes": [
    {"id": "node_id", "label": "Human Readable Name", "source_file": "relative/path", "community": 0, "source_location": "file:line"}
  ],
  "edges": [
    {"source": "node_id", "target": "node_id", "relation": "calls|references|...", "confidence": "EXTRACTED|INFERRED|AMBIGUOUS", "confidence_score": 0.9}
  ],
  "communities": {
    "0": {"label": "Community Name", "node_ids": ["id1", "id2"]}
  },
  "traversal": {"mode": "bfs|dfs", "depth": 3, "budget_used": 1200}
}

• If JSON parses successfully: accumulate nodes, edges, and communities.
• If parsing fails: log warning "Graphify call N failed — skipping." Continue with next call.
• If ALL calls fail: continue to Phase 4 with vault-only content. Never block.

3-G.3 Deduplicate and blend

1. Deduplicate nodes by id, edges by source+target+relation
2. Resolve graphify nodes to vault .md files (by label or source_file)
3. Merge with the vault entities already collected in Phase 2
4. Respect the 15-entity total limit

3-G.4 Check for remote content need

If graphify results reveal additional external URLs that appear relevant to the question and aren't ingested in the vault → escalate to Phase 3-T before proceeding.

---

Phase 3-T — Teach Delegation

Execute when the self-assessment determines needs_remote_content, or when Phase 3-G.4 identifies uningested URLs.

3-T.1 Identify URLs to ingest

From the vault entities read in Phase 2 (and optionally Phase 3-G), identify external URLs that appear directly relevant to answering the question:

• Confluence URLs (containing confluence or atlassian.net)
• Google Docs URLs (containing docs.google.com)
• GitHub URLs (containing github.com)

Limit: 2 URLs per /bedrock:ask invocation. If more than 2 relevant URLs exist, prioritize those most directly related to the question.

3-T.2 Invoke /bedrock:teach

For each URL, invoke /bedrock:teach via the Skill tool:

/bedrock:teach <URL>

Context: Ingesting to answer the question: "<original question>"

IMPORTANT:

• Invoke via the Skill tool — same delegation pattern as teach → preserve
• /teach handles its own flow: fetch content, extract entities, present to user for confirmation, delegate to /preserve
• /ask waits for /teach to complete

3-T.3 Re-read newly created entities

After /teach completes successfully:

1. Search the vault for entities that were just created or updated (based on /teach's output)
2. Read these new entities (frontmatter + body)
3. Add them to the working set of vault entities for response composition

3-T.4 Best-effort fallback

If /teach fails or the user declines the confirmation:

• Log: "Teach delegation for did not complete. Continuing with available content."
• Continue to Phase 4 with whatever content is available
• Never block the response because of a failed teach delegation

---

Phase 4 — Prioritize by Recency

4.1 Identify entities with explicit dates

For discussions and topics, extract the date from the filename:

• Pattern YYYY-MM-DD-slug.md → full date (e.g.: 2026-04-02)
• Pattern YYYY-MM-slug.md → partial date, assume day 01 (e.g.: 2026-04-01)

For consolidated entities (actors, people, teams, projects):

• Treat as equally up-to-date — do not apply temporal ranking
• Trust that content is up-to-date via /bedrock:preserve and /bedrock:compress

4.2 Sort by recency

When the response involves multiple dated discussions or topics:

• Sort by date descending (most recent first)
• If the question is explicitly about something recent ("what happened lately", "latest decisions"), limit to entities from the last 30 days
• If the question is about history ("what happened with X over time"), include all dates but present chronologically (most recent first)

---

Phase 5 — Respond to the User

5.1 Compose the response

Build the response following these rules:

1. Language: Use the vault's configured language. Technical terms in English are accepted (PCI DSS, API, EKS, etc.)

2. Response structure:

   • Open with a direct answer to the question (1-3 sentences)
   • If necessary, expand with details organized by topic
   • Use headers (##, ###) if the response is long (>5 paragraphs)
   • Use tables when the information is comparative or inventory-like

3. Entity citations:

   • Cite ALL consulted entities as wikilinks: [[entity-name]]
   • Use bare wikilinks (never [[dir/entity-name]])
   • Group citations at the end if there are many, or inline when natural

4. Escalation transparency:

   • If graphify was used, note: "I consulted the knowledge graph for deeper context."
   • If /teach was invoked, note: "I ingested [source] into the vault to answer this question."
   • If vault-only was sufficient, no special note needed

5. When nothing is found:

   • State explicitly: "I didn't find information about [X] in the vault."
   • If relevant, suggest: "You can use /bedrock:teach <URL> to ingest a source about this topic."
   • NEVER fabricate information. Only respond with what was found.

6. Response prioritization (Zettelkasten hierarchy): When composing the response, apply weight by Zettelkasten role:

   • Permanent notes (actors, people, teams) — maximum weight, consolidated information. Present as current facts.
   • Bridge notes (topics, discussions) — high weight, contextualized information. Most recent discussions/topics first.
   • Index notes (projects) — medium weight, organizational reference. Point to where the detail is.
   • Fleeting notes — low weight, unconsolidated information. ALWAYS flag with disclaimer: (source: fleeting note — unconsolidated information)
   • If there is conflicting information between sources, point out the discrepancy.

7. Fleeting note promotion detection (criterion 3: active relevance): When a fleeting note is referenced in the response because it is relevant to the query:

   • Check if it meets promotion criteria (see <base_dir>/../../entities/fleeting.md):
     • Critical mass (>3 paragraphs with sources)
     • Corroboration (confirmed by an existing permanent)
   • If any criterion is met, add at the end of the response: > [!info] Promotion suggested: [[fleeting-note-name]] can be promoted to permanent/bridge
   • /bedrock:ask does NOT promote automatically — it only flags. Promotion happens when /bedrock:preserve is invoked with the instruction to promote.

5.2 Post-response suggestions

When appropriate, suggest actions to the user:

• If information is outdated: "The vault may be outdated about [X]. Consider running /bedrock:teach <source> to update."
• If the question revealed gaps: "I didn't find [Y] in the vault. If you have this information, you can use /bedrock:preserve to record it."
• If the question is complex and the response incomplete: "For a more complete view, you may also want to run /bedrock:teach <URL> to ingest additional sources."

---

Critical Rules

Rule	Detail
Vault-first principle	Phase 2 ALWAYS runs before any escalation. Read vault content first, decide later. Never skip Phase 2.
LLM self-assessment	The decision to escalate is made by the LLM after reading vault content (Phase 3.1), not by a heuristic rule table. Use the guidance provided, but the LLM makes the final call.
Escalation priority	When multiple outcomes apply: needs_remote_content > needs_graphify > vault_sufficient. Internalize first, then analyze.
No direct writes	/ask NEVER writes, edits, or deletes files directly. All writes are delegated through /bedrock:teach → /bedrock:preserve.
Teach delegation via Skill tool	Invoke /bedrock:teach via the Skill tool. /teach owns its confirmation gate. /ask cannot bypass it.
Graphify via Skill tool	Invoke /graphify via the Skill tool — NEVER call the Python API directly.
Max graphify calls	Read query.max_graphify_calls from .bedrock/config.json (default: 3, valid range: 1–5). Only consumed when graphify is actually invoked.
Graph unavailable warning	When needs_graphify but graphify-out/graph.json is missing, display > [!warning] callout with /graphify build instruction. Continue with vault-only content.
Best-effort escalation	If graphify fails or teach fails or user declines: continue with available content. NEVER block the response.
Limit of 15 entities	Do not read more than 15 entities total across Phase 2 + Phase 3
Limit of 2 teach URLs	Do not invoke /bedrock:teach for more than 2 URLs per /bedrock:ask invocation
No fabrication	Respond ONLY with information found in the vault or obtained through escalation. Never fabricate data.
Clarification before guessing	If the question is ambiguous, ask for clarification. Do not assume.
Vault language with technical terms in English	Response always in the vault's configured language
Bare wikilinks	[[name]], never [[dir/name]]
Consolidated entities = up-to-date	Actors, people, teams do not need temporal ranking
Dated discussions/topics = prioritize recent	Sort by date in filename (YYYY-MM-DD)
Sensitive data	NEVER display credentials, tokens, PANs, CVVs found in the vault
Fleeting notes with disclaimer	ALWAYS flag information from fleeting notes with (source: fleeting note — unconsolidated information)
Promotion as side-effect	When a relevant fleeting note meets promotion criteria, flag with callout. Do NOT promote automatically.
Weight hierarchy	permanent > bridge > index > fleeting. Use as guideline, not mathematical formula.
Vault resolution first	Resolve VAULT_PATH before any file operation — never assume CWD is the vault
All entity paths use <VAULT_PATH>/ prefix	<VAULT_PATH>/actors/, not actors/