Intent sync keeps agentic coding specs alive. It compares what was learned in a session against repo-local intent artifacts, then proposes the smallest safe updates before writing anything.
Scope
- Repo-local intent artifacts only.
- User approval is required before file writes unless the user explicitly asks for automatic updates.
- Facts, inferences, and recommendations must be labeled separately.
- Lifecycle updates are recorded as new raw entries; old entries are not mutated.
Inputs
- Current conversation or session summary.
- Relevant raw entries and recall context when a vault exists.
- Candidate intent artifacts from:
python <this-skill>/scripts/intent_targets.py --cwd "$PWD"
Output Before Writes
Return a concise proposal:
## Intent Sync Proposal
### Mismatches
- **Fact**: ...
- **Inference**: ...
- **Recommendation**: ...
### Proposed File Updates
- `DESIGN.md`: ...
- `AGENTS.md`: ...
### Minimal Diff Draft
```diff
...
Lifecycle Signals
- Open question -> answered/promoted_to_decision
- Risk -> mitigated/accepted
- Decision -> revised/superseded
Do not rewrite files outside the requested scope. If there is not enough
evidence to update a doc, say so and leave the doc unchanged.
## Approved Write Mode
When the user approves a specific update:
1. Edit only the approved target files.
2. Append a raw `decision` or `risk` entry when vault is available and the
change carries future roadmap value.
3. If a durable artifact was updated and artifact index exists, update or add an
artifact index entry.
## Evidence Rules
- Facts must cite raw entries, current files, or user-provided session context.
- Inferences must use hedging language.
- Recommendations must be framed as proposed changes, not established facts.
- Missing vault data is acceptable; use only the conversation and repo files in
that case.