@neonwatty/limner 0.1.6 → 0.1.8

package/skills/limner/SKILL.md

@@ -17,13 +17,24 @@ Use `limner loop` for Ralph Loop-style polishing across one or more trajectories
 
 1. Start with `limner loop start --mode <mode> --target <name> --name <human-name> --max-iterations <count>`.
 2. Compare with `limner loop compare --trajectory <trajectory-id>`.
-3. Read the generated comparison prompt and response target.
-4. Write or validate the agent response.
-5. Check state with `limner loop status --trajectory <trajectory-id>`.
-6. Move to the next scoped fix with `limner loop next --trajectory <trajectory-id>`.
-7. Close with `limner loop close --trajectory <trajectory-id>`.
+3. Read the generated comparison prompt.
+4. Submit the agent response with `limner loop response submit --trajectory <trajectory-id> --from-run <run-id> --file ./response.json`.
+5. After validation, prefer `limner loop task --trajectory <trajectory-id> --executor subagent`.
+6. Record `limner loop action start --trajectory <trajectory-id> --from-run <run-id> --kind polish --executor subagent --summary "<short edit intent>"` before edits.
+7. Make one scoped edit from the task brief.
+8. Record `limner loop action complete --trajectory <trajectory-id> --action <action-id> --executor subagent --summary "<what changed>" --files "<paths>"` after edits, then rerun `limner loop compare --trajectory <trajectory-id>`.
+9. For comparison-only smoke runs with no intended edit, record `limner loop action skip --trajectory <trajectory-id> --from-run <run-id> --summary "Comparison smoke only; no edit intended"`.
+10. Check state with `limner loop status --trajectory <trajectory-id>`.
+11. Move to the next scoped fix with `limner loop next --trajectory <trajectory-id>`.
+12. Close with `limner loop close --trajectory <trajectory-id>`.
 
-Every meaningful loop interaction writes a ledger event. Use `--feedback "<short note>"` for a 255-character `agentFeedback` note about improving the current process. Use `limner ledger export <trajectory-id> --format markdown` to hand a compact trajectory history to another agent.
+The intended edit loop is `loop compare -> loop response submit -> loop task --executor subagent -> loop action start -> edit -> loop action complete -> loop compare`. Every meaningful loop interaction writes a ledger event. Agent responses are submitted to local SQLite with the full JSON body, hash, validation status, and freshness. Keep every action `--summary` under 255 characters. Use `--feedback "<short note>"` for a 255-character `agentFeedback` note about improving the current process. Use `limner ledger export <trajectory-id> --format markdown` to hand a compact trajectory history to another agent.
+
+Use `limner loop task --trajectory <trajectory-id> --executor subagent --format json` when another tool needs machine-readable task data.
+
+Limner records executor intent and action claims; it cannot prove an external orchestrator actually used a subagent.
+
+Agents can discover the current command surface with `limner --help`, `limner loop --help`, `limner loop response --help`, `limner loop task --help`, and `limner ledger --help`.
 
 ## Image To Reference
 
@@ -33,32 +44,22 @@ Every meaningful loop interaction writes a ledger event. Use `--feedback "<short
 4. Edit `reference/index.html` and `reference/styles.css`.
 5. Run `limner compare image-reference --target <name>`.
 6. Read `captures/image-reference/agent-comparison/agent-prompt.codex.md`.
-7. Write `captures/image-reference/agent-comparison/agent-response.json`.
-8. Rerun `limner compare image-reference --target <name>` so Limner validates `image-comparison.json` and `comparison-summary.json`.
+7. Submit JSON with `limner loop response submit --trajectory <trajectory-id> --from-run <run-id> --file ./response.json`.
+8. Use the validated `image-comparison.json` and `comparison-summary.json`.
 9. Use the validated next-iteration guidance for one scoped mockup fix.
 
-Optional structured spec layer:
-
-- Run `limner compare image-reference --target <name> --spec`.
-- Edit `contract/visual-spec-instructions.md` to customize the agent prompt for this target, or pass `--spec-instructions <path>` for a shared policy file.
-- Use `captures/image-reference/spec/agent-prompt.codex.md` for Codex or `captures/image-reference/spec/agent-prompt.claude.md` for Claude.
-- Use the generated `captures/image-reference/spec/agent-response.schema.json` for strict structured output when the agent supports it.
-- The agent should inspect the ideal image and reference screenshot separately. The side-by-side is comparison context, not the image to parse.
-- Have the agent write `captures/image-reference/spec/agent-response.json`.
-- Rerun the same command so Limner can validate the response and emit the spec and diff JSON artifacts.
-
 ## Reference To Implementation
 
 1. Confirm the reference HTML is approved.
 2. Ensure every important region has both `referenceSelector` and an implementation selector in `appSelector`.
 3. Run `limner compare reference-implementation --target <name> --url <implementation-url>`.
 4. Read `captures/reference-implementation/agent-comparison/agent-prompt.codex.md`.
-5. Write `captures/reference-implementation/agent-comparison/agent-response.json`.
-6. Rerun the same command so Limner validates image, structure, and summary comparison artifacts.
+5. Submit JSON with `limner loop response submit --trajectory <trajectory-id> --from-run <run-id> --file ./response.json`.
+6. Use the validated image, structure, and summary comparison artifacts.
 7. Use the validated next-iteration guidance for one scoped implementation fix.
 
 ## Constraints
 
-- Limner does not call an AI vision model itself. Agent comparison and the `--spec` workflow require an external agent.
+- Limner does not call an AI vision model itself. Agent comparison requires an external agent.
 - Limner does not make final pass/fail judgments.
 - Limner logs only local JSONL events under `.limner/runs/`.

package/templates/target/AGENT_GUIDE.md

@@ -11,8 +11,8 @@ Prompt profiles: `ideal-to-mockup` for Loop 1, and `mockup-to-implementation` fo
 3. Edit `reference/index.html` and `reference/styles.css` until the reference is a faithful HTML facsimile.
 4. Run `limner compare image-reference --target <target>`.
 5. Read `captures/image-reference/agent-comparison/agent-prompt.codex.md`.
-6. Write `captures/image-reference/agent-comparison/agent-response.json`.
-7. Rerun the comparison and use the validated next-iteration guidance for one scoped mockup fix.
+6. Submit JSON with `limner loop response submit --trajectory <trajectory-id> --from-run <run-id> --file ./response.json`.
+7. Use the validated next-iteration guidance for one scoped mockup fix.
 
 ## Loop 2: Reference To Implementation
 
@@ -20,8 +20,8 @@ Prompt profiles: `ideal-to-mockup` for Loop 1, and `mockup-to-implementation` fo
 2. Add implementation selectors to `contract/regions.json` using `appSelector`.
 3. Run `limner compare reference-implementation --target <target> --url <implementation-url>`.
 4. Read `captures/reference-implementation/agent-comparison/agent-prompt.codex.md`.
-5. Write `captures/reference-implementation/agent-comparison/agent-response.json`.
-6. Rerun the comparison and use the validated next-iteration guidance for one scoped implementation fix.
+5. Submit JSON with `limner loop response submit --trajectory <trajectory-id> --from-run <run-id> --file ./response.json`.
+6. Use the validated next-iteration guidance for one scoped implementation fix.
 
 ## Rule