| name | capture |
| description | Use when inventorying conversation state into durable artifacts, routing shovel-ready items toward exec, or deep-processing a brain dump. |
| skill_type | workflow |
| category | lifecycle |
| output_template | hud |
| triggers | ["capture","dump"] |
/capture - Shape to Plan
Inventory conversation state, write durable artifacts, and route each entity to
its next lifecycle owner. /dump is capture --deep.
Work Link
Lifecycle lane: Shape -> Plan
Entity movement: memory -> captured | proposed | blocked
Workstream: resolve active workstream before writing capture artifacts
Upstream: conversation, /dump, /work
Downstream: /prior-art, /propose, /exec
Router: /work
HUD: end with 🧬 {ws} ⚡{exec_count} 📥{capture_count} ⏸️{paused_count} ✅{done_count} | 🚦{gate}={score} | ⏭️ {next} | 🔁{loop_state}
Canon Write Gate
If the user asks "what would you edit", "show me", "audit", "plan", "pass2",
"recommend", or "how would you", do not mutate canon files. Render the ledger
and proposed changes only. Write artifacts only when the user explicitly asks
to apply, edit, patch, emit, or create a durable capture.
Protocol
steps:
- id: inventory
action: Inventory ideas, decisions, follow-ups, artifacts, blockers, promises, and proof/QA risks.
validation: "Every item has a topic and provisional lifecycle state."
on_failure: "If the thread is too large, bucket by topic and continue."
- id: classify
action: Compile user/source prose into agent-operational intent, state, and route.
validation: "Each item has compiled_intent, route_state, and next_route."
on_failure: "Unknown state stays in memory and becomes the first output bucket."
- id: prior_art_probe
action: For non-trivial items, search existing work before creating new files.
validation: "Existing home found, or 'no prior art found' is stated."
Routes
| Item | Route |
|---|
| Constraint, invariant, gate, policy | DNA-backed task via /propose |
| Brief/document-artifact framing | /brief -> /interview --auto -> .lev/pm/designs/ then /propose |
| Design-grade framing | /interview --auto -> .lev/pm/designs/ then /propose |
| Runtime, agentic, promotion, cleanup, fallback, or boundary-risk item | /interview or /propose with qa_seed |
| Execution-ready task with verifier and write scope | /exec |
| Provenance or duplicate check needed | /prior-art |
| Ambiguous intent or boundary | /interview |
| Workstream identity missing | /ws |
Deep Mode
/capture --deep and /dump add a convergence loop:
deep_capture:
repeat_until: "all writable items have stored fidelity >= in-memory fidelity"
loop:
- inventory by topic
- prior-art per topic
- write or update durable route
- score fidelity
- enrich weak captures
circuit_breaker: "3 iterations without fidelity improvement"
Fidelity Gate
fidelity = 0.30*detail_preservation
+ 0.25*relationship_preservation
+ 0.20*source_attribution
+ 0.15*neighbor_context
+ 0.10*actionability
fidelity >= 0.8 can advance to /propose or /exec; lower scores stay in
capture, prior-art, or interview.
Lifecycle Ledger Contract
The lifecycle ledger is the core visibility primitive at lifecycle trigger
points. /capture, /capture --deep, and /dump all render the same table
before claiming route advancement.
| Field | Meaning |
|---|
intent_id | Stable row id for this lifecycle turn |
topic | Short label for the source material |
compiled_intent | Agent-operational interpretation, not raw user prose |
current_location | disk, memory, both, external, or unknown |
artifact_ref | Path or durable id; none if only in memory |
route_state | in_memory, captured, proposed, execution_ready, blocked, done, or rejected |
fidelity | Preservation score plus reason if below 0.8 |
next_route | /interview, /prior-art, /propose, /exec, /close, or none |
blocker | null or the reason this row cannot advance |
No route can advance if a row has current_location: memory without a blocker
or planned destination. This prevents hidden in-chat state from being mistaken
for durable work.
Durable Artifact
Write capture reports near the workstream:
path: .lev/pm/workstreams/<ws-id>/captures/YYYYMMDD-capture-sweep.yaml
schema:
date: YYYY-MM-DD
workstream: <ws-id>
ledger:
- intent_id: "<stable row id>"
topic: "<summary>"
compiled_intent: "<agent-operational interpretation>"
current_location: disk | memory | both | external | unknown
artifact_ref: "<path, durable id, or none>"
route_state: in_memory | captured | proposed | execution_ready | blocked | done | rejected
fidelity: 0.0
next_route: "<skill/path/none>"
blocker: "<reason or null>"
items:
- topic: "<summary>"
stage: captured | proposed
[]
[]
Markdown capture reports are projections only when useful.
Output
## /capture results
Ledger
| ID | Topic | Compiled Intent | Current Location | Artifact | Route State | Fidelity | Next Route | Blocker |
|---|
| {intent_id} | {topic} | {compiled_intent} | {current_location} | {artifact_ref} | {route_state} | {fidelity_pct} | {next_route} | {blocker} |
In Memory
{items_that_could_not_be_written}
Blocked
{blocked_items}
Crystallizing
{transitioning_items}
Stats
- Total inventoried: {n}
- In memory captured: {n}
- Transitioning: {n}
- Blocked: {n}
- Ledger rows: {n}
Red Flags
- "I'll show the list first and save after approval."
- "This follow-up is obvious enough to leave in chat."
- "Done items should be listed again."
- "Prior art can wait until proposal."
- "The table is optional because I wrote the artifact."
- "A routed item can stay only in chat memory."
- "Copying user prose is the same as compiling intent."
- "What would you edit means patch it now."
Related
/work routes lifecycle lanes./dump is a focused alias for /capture --deep./brief turns conversation context into a lifecycle artifact via /interview --auto./prior-art finds existing homes./propose turns captured/aligned items into task DNA./exec runs shovel-ready items.
