sync

/sync

Read $AI_INFRA_DIR/extract-buffer.md, distribute each signal to the appropriate target files, then clear the buffer.

Path configuration: $AI_INFRA_DIR must be set before use. This is the directory where all Contour data files are stored.

• Default recommended path: ~/.claude/contour/
• Set via environment variable, or replace this placeholder with your absolute path.

---

Reference Files

Before executing, read the following reference files for format specifications:

• references/domain-state-structure.md — Domain State table format and update rules

---

File References

File	Term	Role	Access
extract-buffer.md	Extract Buffer	Temporary buffer: read then clear	Read + Clear
{user}-{domain}.md	Domain State	Current cognitive state for the domain	Read + Write
{user}-core.md	Core Profile	Personality-level traits, thinking patterns, core preferences	Read only (never write)

MVP scope: Single domain only. The routing principle below applies regardless: route signals by content domain, not by the source field (source indicates where the signal was captured, not its domain).

Multi-domain support (additional Domain State file pairs) is a Phase 2 extension.

---

Constraints

• Only read Extract Buffer, Domain State, Core Profile
• Never write to Core Profile — /sync does not modify core.md. If a signal may warrant a Core Profile update, surface it as a [core-candidate] in the sync report for the user to decide immediately.
• Do not re-scan conversation history — all signals come from Extract Buffer
• Do not invent signals not present in Extract Buffer
• Preserve all existing content in Domain State — only append or update specific entries

---

Pre-flight Check

Before processing, verify the state of Extract Buffer:

• Extract Buffer file does not exist → Inform the user: "extract-buffer.md not found. Check your path configuration." Stop.
• Extract Buffer file exists but is empty → Inform the user: "Buffer is empty, nothing to sync." Stop.
• Extract Buffer file exists with content → Proceed.

---

Step-by-Step Execution

Step 1: Read Files

Read the reference files listed above. Read Extract Buffer in full. Read Domain State in full. Read Core Profile in full (for deduplication and comparison only — never write to Core Profile).

Step 2: Process Each Signal

For every signal in Extract Buffer, determine its type and execute the corresponding routing rule.

[cognition] → Domain State

1. Check Domain State for an existing entry on the same knowledge point.
   • Same knowledge point = same concept/tool/process, regardless of wording differences (e.g., "git rebase vs merge" and "difference between rebase and merge" are the same)
   • If an existing entry has the same cognitive status, this signal is a duplicate → skip Domain State update, record as duplicate in report
   • If an existing entry has a lower cognitive status (e.g., partial → mastered), move the ✓ in Domain State to the higher column per domain-state-structure.md, update the date
   • If an existing entry has a higher cognitive status (e.g., mastered → partial), this is a cognitive regression → move the ✓ in Domain State to the lower column, update the date
   • If no existing entry, add a new row to Domain State

[thinking] → Sync report only

1. Include in the sync report under "Thinking patterns observed" — do not write to any file.

2. Check Core Profile for relevance:

   • If this thinking pattern is not represented in Core Profile and appears to be a persistent trait → surface as [core-candidate] in the sync report
   • If this thinking pattern is already captured in Core Profile → note as "already tracked" in report
   • If this thinking pattern contradicts something in Core Profile → surface as [core-candidate] in report with a contradiction note

[preference] → Domain State and/or [core-candidate]

1. Determine scope:

   • If the preference is domain-specific (e.g., "when explaining code, use terminology") → target Domain State's communication rules section
   • If the preference is cross-domain (e.g., "don't use metaphors", "be direct") → do not write to Core Profile; instead surface as [core-candidate] in the sync report
   • If ambiguous, default to Domain State (narrower scope is safer)

2. For Domain State-targeted preferences, check for existing similar preferences in Domain State:

   • If a semantically equivalent preference already exists → skip (duplicate)
   • If the new preference contradicts an existing one → replace the old one with the new one
   • If no existing equivalent → add to Domain State

Step 3: Clean Up Domain State

After all signals are processed:

• If Domain State has grown noticeably, review for entries that can be compressed or consolidated
• Items marked mastered for more than 30 days with no further status changes can be removed from Domain State
• Report to the user if any cleanup was performed

Step 4: Clear Extract Buffer

After all signals have been distributed successfully, clear the content of Extract Buffer (keep the file, empty its content).

Step 5: Report

Sync complete ({date}):
- Processed: {n} signals from {m} extract blocks
- Domain State updated: {list of changes, or "no changes"}
- Thinking patterns observed: {list of summaries, or "none"}
- Core Profile candidates (your decision needed): {list, or "none"}
- Duplicates skipped: {number}
- Buffer cleared

For each [core-candidate] surfaced, ask the user directly:

"This pattern/preference isn't in your Core Profile yet: [{description}]. Add it? (yes / no / edit)"
If yes → instruct user to manually add to their {user}-core.md. Do not write it yourself.

After reporting, remind the user:

This session is now complete. Do not run /contour:extract in this session — the sync conversation would produce echo signals.

---

Deduplication Rules

Deduplication is semantic, not lexical. Two entries refer to the same thing if:

1. They describe the same concept, tool, or process (e.g., "environment variables" and "env vars" and ".env files" are the same if they refer to the same knowledge point)
2. They describe the same thinking pattern (e.g., "validate before expanding" and "don't expand until the core is proven" are the same pattern)
3. They describe the same preference (e.g., "use terminology" and "don't dumb down technical terms" express the same preference)

When in doubt, keep both entries — false deduplication (losing a real signal) is worse than mild redundancy (which can be cleaned up manually).

---

Edge Cases

• Multiple extract blocks in Extract Buffer: Process them chronologically. Later signals may override earlier ones for the same knowledge point.
• Signals from different workspaces: Route by content domain, not by source workspace. A coding cognition signal captured in Content_Creator still goes to the coder domain files.
• Conflicting signals in the same sync: If two signals in Extract Buffer contradict each other (e.g., one says "partial", another says "mastered" for the same knowledge point), use the later one (it reflects more recent state).
• Missing target files: If Domain State doesn't exist yet, create it with the appropriate structure per the reference files before writing.