project-init

Project Init

Overview

All subagent dispatches use disk-mediated dispatch. See shared/dispatch-convention.md for the full protocol.

Eliminate cold-start penalty by proactively mapping the current repo and its neighborhood. Instead of re-discovering the codebase during every first task, project-init builds structural context upfront so that build, design, and debugging skills start informed.

Invocation: User runs /project-init. Not auto-triggered.

Two tiers:

• Tier 1 — Single repo deep scan. Fan-out partition explorers, fan-in to cartographer format.
• Tier 2 — Cross-repo discovery. Scan sibling repos for topology and dependency relationships.

Output:

• Tier 1 writes to cartographer's existing data structures (memory/cartographer/)
• Tier 2 writes to memory/topology/

Coverage distinction: All output is tagged <!-- project-init:structural -->, marking it as breadth-first structural mapping. This is distinct from task-verified content produced by cartographer record mode during real work. Task-verified content is always preserved over structural content.

Announce at start: "I'm using the project-init skill to map this codebase and its neighborhood."

---

Pre-flight: Scope Estimation

Before dispatching any agents, the orchestrator estimates work scope:

1. Count top-level source directories by scanning for files with recognized extensions: .ts, .js, .py, .go, .rs, .java, .cs, .rb, .swift, .kt, .c, .cpp, .h
2. Count manifest files (package.json, go.mod, Cargo.toml, pyproject.toml, docker-compose.yml) and sibling repos (git repos in parent directory)

Present the estimate to the user:

"Found N source directories and M sibling repos. This will dispatch ~X agents. Proceed?"

User options:

• Approve — run both tiers
• Skip Tier 2 — run Tier 1 only (single repo scan)
• Narrow scope — user specifies which directories or repos to include

Wait for user confirmation before proceeding. Do not dispatch agents without approval.

---

Tier 1: Single Repo Deep Scan

Step 0: Cleanup

Delete /tmp/crucible-project-init/ if it exists from a prior run, then recreate it. This ensures a clean workspace for temp files.

rm -rf /tmp/crucible-project-init && mkdir -p /tmp/crucible-project-init

Step 1: Detect Project Structure

Scan the repository to determine:

• Source directories — top-level directories containing files with recognized source extensions (.ts, .js, .py, .go, .rs, .java, .cs, .rb, .swift, .kt, .c, .cpp, .h)
• Ecosystem — inferred from manifest files:
  • package.json → Node/JavaScript/TypeScript
  • go.mod → Go
  • Cargo.toml → Rust
  • pyproject.toml / requirements.txt → Python
  • pom.xml / build.gradle → Java/Kotlin
  • *.csproj / *.sln → C#/.NET
  • Gemfile → Ruby
  • Package.swift → Swift
• Config/doc directories — directories with no source files (docs, config, assets)

Small Repo Shortcut

If the entire repository has fewer than 20 source files (across all directories), skip the partition/fan-out/fan-in pipeline. Instead, dispatch a single Partition Explorer for the repo root and pass its output directly to the Init Recorder as a single partition report. This avoids unnecessary overhead for small codebases.

Step 2: Partition

Split the repository into partitions for parallel exploration:

1. One partition per top-level source directory (e.g., src/, lib/, pkg/, cmd/)
2. Small directories (<20 source files) — single explorer handles the entire directory
3. Doc/config-only directories — lightweight explorer (reports "No source modules. Purpose: docs/config/assets")
4. Large directories (50+ source files) — sub-partition by next directory level. For example, src/ with 80 files becomes src/auth/, src/api/, src/models/, etc.
   • Max 1 level of sub-partitioning. If a sub-partition still has 50+ files, the explorer handles it as a large partition (it will apply its own internal triage)

Step 3: Fan-out — Partition Explorers

Dispatch parallel Explore subagents, one per partition:

Agent tool (subagent_type: Explore, model: sonnet)

Use the prompt template at ./partition-explorer-prompt.md. Fill in the template variables:

• [partition name] — the partition directory name
• [Partition root directory path] — absolute path to the partition
• [Source file extensions detected in this partition] — which extensions were found
• [Project ecosystem context] — ecosystem info from Step 1

When each explorer returns: The orchestrator captures the return value and writes it to /tmp/crucible-project-init/<partition-name>.md. The explorer itself does not write files — the orchestrator does.

Step 4: Fan-in — Init Recorder

Dispatch the Init Recorder to merge all partition reports into cartographer format:

Task tool (general-purpose, model: sonnet)

Use the prompt template at ./init-recorder-prompt.md. Fill in the template variables:

• [N] — number of partition reports
• [File paths to partition exploration reports] — paths to temp files from Step 3
• [Existing cartographer data] — read and paste existing memory/cartographer/map.md, conventions.md, landmines.md if they exist, or say "No prior cartographer data."
• [Project name and ecosystem] — from Step 1
• [Output directory] — the project's memory/cartographer/ path

Batching for large repos (6+ partitions): When there are 6 or more partition reports:

1. Group reports into batches of 5
2. Dispatch one Init Recorder per batch with "batch mode" in the description — each writes its merged output to /tmp/crucible-project-init/batch-N.md in explorer format (the same ## Modules Found, ## Conventions Observed, etc. sections used by partition explorers), NOT cartographer format. This is a consolidation pass, not a formatting pass. Set the Output Directory to /tmp/crucible-project-init/batch-N.md.
3. Dispatch a final Init Recorder (without "batch mode") that receives the batch output file paths and produces the definitive cartographer files. The final recorder treats batch files exactly like partition reports — same input format, same processing steps.
4. The final pass handles deduplication and conflict resolution across batches

The orchestrator passes file paths, not content for partition reports to the Init Recorder — the recorder reads the partition reports itself. Existing cartographer data (for re-invocation merge) is included directly in the dispatch file since it's small and needed for merge decisions.

Step 5: Validation Gate

After the Init Recorder completes, run a three-way check:

(a) Partition completeness — Verify each partition explorer returned a non-empty result. Check that temp files exist at /tmp/crucible-project-init/<partition-name>.md and contain non-trivial content (more than 3 lines). Also check for the completion sentinel: <!-- partition-explorer:complete --> means full scan, <!-- partition-explorer:partial --> means some directories were unscanned (flag these in map.md Unmapped Areas).

(b) Map representation — Verify every partition that returned results is REPRESENTED in map.md — either as individual modules OR as a collapsed group in "Other" (collapsed partitions count as represented). Unrepresented partitions trigger focused re-recording: dispatch the Init Recorder again with just the missing partition reports.

(c) Module field completeness — Verify module entries have required fields populated: Path, Responsibility, Key Components must all be non-empty. Also verify that a modules/<name>.md file exists for each module listed in map.md with Mapped Detail = Yes.

Partitions that returned empty results are flagged as "unmapped" in map.md under the Unmapped Areas section.

Step 6: Context Scan

Read the following files if they exist:

• README.md — project purpose, setup instructions
• CONTRIBUTING.md — contribution guidelines, review process
• CLAUDE.md — existing agent instructions

This is a direct read by the orchestrator, not a subagent dispatch. The orchestrator uses the extracted context for two purposes:

1. CLAUDE.md proposal filtering (Step 7) — avoid proposing content that duplicates existing CLAUDE.md
2. Memory creation — if README reveals project purpose, team conventions, or setup requirements not already in memory, save as project type memories via the auto-memory system

Step 7: CLAUDE.md Proposal (Non-blocking)

If the Init Recorder produced /tmp/crucible-project-init/claude-md-proposal.md:

1. Read the existing project CLAUDE.md (if any) to identify already-configured content
2. Filter out proposals that duplicate existing CLAUDE.md content
3. Write the filtered proposal to /tmp/crucible-project-init/claude-md-proposal-filtered.md — this survives context compaction during Tier 2
4. Do not present proposals yet — continue to Tier 2. Proposals are presented at the END of the full run (after both tiers complete)

This step does NOT block Tier 2 — the pipeline continues autonomously.

Size Caps

File	Target	Hard Cap
map.md	140 lines	200 lines
conventions.md	105 lines	150 lines
landmines.md	70 lines	100 lines
modules/<name>.md	70 lines	100 lines

Target 70% of caps — leave room for task-verified additions by cartographer record mode.

Large Monorepo Triage

When the repository has many top-level source directories:

• 20+ partitions: Warn user: "Large monorepo detected (N source directories). Recommend narrowing scope to specific areas." Offer to scan a subset.
• Collapsed modules: If the unified module count exceeds the map.md cap, collapse low-file-count modules into an "Other" row with count: | Other | various | 12 single-file modules | No |
• Sub-partitioning cap: Never sub-partition more than 1 level deep. Large sub-partitions are handled by the explorer's own triage logic.

Orchestrator writes Tier 1 results to disk before proceeding to Tier 2.

---

Tier 2: Cross-Repo Discovery

Step 0: Permission Probe

Before scanning neighbors, verify filesystem access:

1. Attempt to list the parent directory (../)
2. Attempt to read a file from a detected sibling repo

If access is denied, emit a clear skip message and end Tier 2:

"Cross-repo discovery requires filesystem access to the parent directory. Skipping Tier 2."

Step 1: Manifest Parsing

Parse supported manifest formats for cross-repo references:

Format	What to Extract
package.json	dependencies, devDependencies — look for file:../ or workspace references
go.mod	require directives — look for local replace directives pointing to siblings
Cargo.toml	[dependencies] — look for path = "../" references
pyproject.toml	[project.dependencies] — look for local path references
docker-compose.yml	services — look for build paths pointing to siblings, shared networks/volumes

Note detected-but-unparsed formats (e.g., pom.xml found but not parsed) so the user knows what was skipped.

Step 2: Local Sibling Detection

Scan the parent directory for git repos:

• List directories in ../
• Check each for .git/ directory
• Cross-reference with manifest references from Step 1
• Classify each sibling:
  • Manifest-referenced — found in a manifest file (pre-selected for scanning)
  • Co-located — git repo in same parent directory, no manifest reference

Step 3: User Confirmation

Present discovered repos to user before scanning:

"Found N sibling repos. M are referenced in manifests (pre-selected). Confirm which to scan:"

• ../auth-service (referenced in docker-compose.yml)
• ../shared-types (referenced in package.json)
• ../unrelated-project (co-located, no reference)

Wait for user confirmation. Do not scan repos the user did not confirm.

Step 4: Lightweight Neighbor Scan

Dispatch parallel Explore subagents, one per confirmed neighbor:

Agent tool (subagent_type: Explore, model: sonnet)

Use the prompt template at ./neighbor-scanner-prompt.md. Fill in the template variables:

• [repo name] — the neighbor repository name
• [Neighbor repo path] — path to the neighbor
• [Connection context] — how it was discovered (manifest reference details)
• [Current repo name and purpose] — from Tier 1 results

Write each result to /tmp/crucible-project-init/neighbors/<repo-name>.md.

Step 5: Relevance Ranking

After all neighbor scans complete, the orchestrator assigns relevance:

Relevance	Criteria	Example
High	Direct dependency — imported, called, or required by current repo	file:../shared-types in package.json
Medium	Shared infrastructure — common services, shared DB, docker-compose links	Both use same Redis instance
Low	Co-located, no detected link — just happens to be in same parent directory	../unrelated-tool

Step 6: Topology Output

Dispatch the Topology Recorder to synthesize neighbor scans:

Task tool (general-purpose, model: sonnet)

Use the prompt template at ./topology-recorder-prompt.md. Fill in the template variables:

• [N] — number of neighbor scans
• [File paths to neighbor scan results] — paths to temp files from Step 4
• [Current repo name, ecosystem, purpose] — from Tier 1
• [Relevance scores] — per-neighbor relevance from Step 5
• [Existing topology data] — read and paste existing memory/topology/topology.md if it exists, or say "No prior topology data."
• [Output directory] — the project's memory/topology/ path

Output Structure

After both tiers complete, the following structure exists:

~/.claude/projects/<project-hash>/memory/
  cartographer/
    map.md                    # Module map (← Tier 1)
    conventions.md            # Codebase patterns (← Tier 1)
    landmines.md              # Non-obvious breakage (← Tier 1)
    modules/
      <name>.md               # Per-module detail (← Tier 1)
  topology/
    topology.md               # Cross-repo dependency map (← Tier 2)
    <neighbor-name>.md        # Per-neighbor detail (← Tier 2)

---

Completion: CLAUDE.md Proposal

After BOTH tiers complete (or after Tier 1 if Tier 2 was skipped), present the CLAUDE.md proposal if one was generated. Read from /tmp/crucible-project-init/claude-md-proposal-filtered.md (the filtered version from Step 7):

"Structural mapping complete. Also generated CLAUDE.md proposals based on what I found. Review below — merge what's useful."

[display proposal content]

User options:

• Accept all — orchestrator appends all proposed content to the project's CLAUDE.md
• Accept selectively — user indicates which sections to keep
• Skip — no changes to CLAUDE.md

The orchestrator appends accepted content to the project's CLAUDE.md (creating the file if it doesn't exist). Appended content is added under a clear heading.

---

Subagent Dispatch Summary

Agent	Model	Dispatch	Prompt Template
Partition Explorer	Sonnet	Agent tool (Explore)	./partition-explorer-prompt.md
Init Recorder	Sonnet	Task tool (general-purpose)	./init-recorder-prompt.md
Neighbor Scanner	Sonnet	Agent tool (Explore)	./neighbor-scanner-prompt.md
Topology Recorder	Sonnet	Task tool (general-purpose)	./topology-recorder-prompt.md

Explorers are dispatched via the Agent tool with the specified subagent_type. Recorders are dispatched via the Task tool (general-purpose). Use the prompt templates verbatim, filling in only the bracketed template variables.

---

Agent Teams Fallback

If agent teams are not available (Agent tool does not support parallel dispatch), fall back to sequential dispatch with a one-time warning:

"Agent teams not available. Running sequentially — this will take longer."

Behavior is unchanged except parallel dispatch becomes sequential. All steps, validation, and output remain the same.

---

Re-invocation Merge Strategy

When project-init is run again on a repo with existing cartographer or topology data:

Existing Content	Action
<!-- project-init:structural --> tagged	Overwrite with fresh scan data
Task-verified (no structural tag)	Preserve — never modify or remove
New modules/neighbors not in prior data	Add with structural tag
Prior modules/neighbors absent from scan	Flag with [STALE?] marker — do not remove
Overflow after merge	Prioritize task-verified content, compress structural

This strategy ensures that knowledge accumulated through real task work is never lost by a re-scan.

---

Context Management

Project-init is context-intensive. Follow these rules to prevent context exhaustion:

1. Tier 1 and Tier 2 are separate phases — complete Tier 1 and write all results to disk before starting Tier 2
2. Explorer outputs go to temp files — the orchestrator writes explorer return values to /tmp/crucible-project-init/ and passes file paths (not content) to recorders
3. Never hold all explorer outputs in orchestrator context — write each to disk as it returns
4. Context pressure at 50% — if the orchestrator reaches 50% context utilization, write accumulated data to disk, report partial progress to the user, and continue with remaining work
5. Batching for large repos — 6+ partition reports are batched through multiple recorder passes (see Step 4)

---

Red Flags

Never:

• Hold all explorer outputs in orchestrator context simultaneously
• Exceed file size caps (200 lines map.md, 150 conventions.md, 100 landmines.md, 100 module files)
• Produce speculative content — record observed facts only
• Scan repos the user didn't confirm in Tier 2
• Skip the permission probe before Tier 2
• Skip the scope estimation or proceed without user approval

Always:

• Write to disk between tiers
• Tag all output with <!-- project-init:structural -->
• Present scope estimate and wait for user confirmation
• Respect the user's scope narrowing choices
• Validate after fan-in (three-way check)
• Preserve task-verified content during re-invocation
• Clean up /tmp/crucible-project-init/ at the start of each run

---

Integration

Required Downstream Change

Cartographer recorder-prompt.md must handle structural tags — when updating files that contain <!-- project-init:structural --> content, the recorder preserves that tag on structural sections and omits it on task-verified additions.

Downstream Consumption

Skills that benefit from project-init output:

Skill	How It Uses project-init Data
crucible:cartographer (consult)	Reads map.md — structural content provides baseline even before any task exploration
crucible:cartographer (load)	Loads modules/*.md into subagent prompts — structural context prevents wrong assumptions
crucible:build	Gets structural awareness from cartographer consult at task start
crucible:design	Knows module boundaries and dependencies before proposing architecture
crucible:debugging	Loads module context and landmines for investigators

Does NOT

• Seed forge (forge learns from agent behavior, not codebase structure)
• Clone remote repos (works only with local filesystem)
• Run tests or install dependencies (read-only scan)
• Modify any source code

Related Skills

• crucible:cartographer — ongoing codebase mapping (project-init bootstraps, cartographer maintains)
• crucible:build — implementation workflow (consumes cartographer data)
• crucible:design — architecture planning (consumes map and topology)
• crucible:debugging — investigation workflow (consumes modules and landmines)

Does not dispatch /recon -- bootstraps the cartographer data that /recon consults. Complementary, not overlapping. See #147 for rationale.

---

Prompt Templates

• ./partition-explorer-prompt.md — Structured exploration per partition for Tier 1
• ./init-recorder-prompt.md — Multi-source fan-in recorder for Tier 1
• ./neighbor-scanner-prompt.md — Lightweight neighbor exploration for Tier 2
• ./topology-recorder-prompt.md — Topology file writer for Tier 2