feat: add data flywheel trainer artifacts

Author: danielfoch

.agent/AGENTS.md

@@ -55,6 +55,9 @@ Daily driver, highest-leverage first:
   in one shot (stage + graduate + render). For rules you already know.
 - `show.py` — one-screen dashboard of brain state: episodes, candidates,
   lessons, failing skills, activity graph.
+- `data_flywheel_export.py` — local export of approved, redacted runs into
+  trace records, context cards, eval cases, training-ready JSONL, and
+  flywheel metrics. It does not train models or call APIs.
 - `list_candidates.py` / `graduate.py` / `reject.py` / `reopen.py` — review
   protocol for patterns the dream cycle has staged.
 - `memory_reflect.py <skill> <action> <outcome>` — log a significant event.

.agent/skills/_index.md

@@ -26,3 +26,11 @@ Pre-deployment verification against a structured checklist.
 Triggers: "deploy", "ship", "release", "go live"
 Constraints: all tests passing, no unresolved TODOs in diff,
 requires human approval for production.
+
+## data-flywheel
+Turns approved, redacted runs into reusable local artifacts: trace records,
+context cards, eval cases, training-ready JSONL, and flywheel metrics.
+Triggers: "data flywheel", "trace to train", "training traces",
+"context cards", "eval cases", "approved runs", "vertical intelligence"
+Constraints: local-only by default; human-approved runs only; redaction required
+before trainable; does not train models.

.agent/skills/_manifest.jsonl

@@ -3,3 +3,4 @@
 {"name":"git-proxy","version":"2026-01-01","triggers":["commit","push","branch","merge","rebase","pull request","PR"],"tools":["bash"],"preconditions":[".git exists"],"constraints":["never force push to main","never force push to protected branches","run tests before push"],"category":"operations"}
 {"name":"debug-investigator","version":"2026-01-01","triggers":["debug","why is this failing","investigate","stack trace","bug"],"tools":["bash","memory_reflect"],"preconditions":[],"constraints":["reproduce before fixing","fix root cause, not symptoms"],"category":"engineering"}
 {"name":"deploy-checklist","version":"2026-01-01","triggers":["deploy","ship","release","go live"],"tools":["bash"],"preconditions":[],"constraints":["all tests passing","no unresolved TODOs in diff","requires human approval for production"],"category":"operations"}
+{"name":"data-flywheel","version":"2026-04-25","triggers":["data flywheel","trace to train","training traces","context cards","eval cases","approved runs","vertical intelligence"],"tools":["bash","git"],"preconditions":[".agent exists"],"constraints":["local-only by default","human-approved runs only","redaction required before trainable","do not train models"],"category":"operations"}

.agent/skills/data-flywheel/SKILL.md

@@ -0,0 +1,112 @@
+---
+name: data-flywheel
+version: 2026-04-25
+triggers: ["data flywheel", "trace to train", "training traces", "context cards", "eval cases", "approved runs", "vertical intelligence"]
+tools: [bash, git]
+preconditions: [".agent exists"]
+constraints: ["local-only by default", "human-approved runs only", "redaction required before trainable", "do not train models"]
+---
+
+# Data Flywheel - approved runs into reusable intelligence
+
+Use this skill when a user wants to turn repeated human-approved agent work
+across Claude Code, Hermes, OpenClaw, Codex, Cursor, or custom `.agent/` loops
+into local artifacts for retrieval, evals, prompt shrinking, and optional
+future open-weight model/adapters.
+
+The flywheel is:
+
+```text
+approved run
+-> redacted trace
+-> context card
+-> eval case
+-> training-ready JSONL
+-> optional downstream SLM/adapter experiment later
+```
+
+This skill creates the harness. It does not train a model.
+
+## Hard Rules
+
+- Use only human-approved runs. Rejected or unknown-review runs can become
+  failure-mode notes, not trainable examples.
+- Redaction must pass before anything is marked trainable.
+- Do not store raw prompts, raw code, client names, addresses, phone numbers,
+  emails, secrets, credentials, or unredacted CRM records.
+- Keep `.agent/flywheel/` private and gitignored unless the user explicitly
+  commits sanitized examples.
+- Stay model-agnostic. Mention model families only as downstream examples.
+
+## Inputs
+
+Default local input:
+
+```text
+.agent/flywheel/approved-runs.jsonl
+```
+
+Each line should be a sanitized run record with:
+
+- `domain`
+- `workflow`
+- `harness`
+- `instruction`
+- `input_redacted`
+- `output_approved`
+- `human_review.status` as `accepted` or `edited`
+- `redaction_status: passed`
+- `pii_level`
+- optional `stable_rules`, `tool_contracts`, `eval_tags`, `failure_modes`
+
+## Export
+
+Run:
+
+```bash
+python3 .agent/tools/data_flywheel_export.py
+```
+
+Outputs go to:
+
+```text
+.agent/flywheel/exports/<YYYY-MM-DD>/
+```
+
+Key outputs:
+
+- `trace-records.jsonl`
+- `training-examples.jsonl`
+- `eval-cases.jsonl`
+- `context-cards/<domain>/<workflow>.md`
+- `context-cards/<domain>/<workflow>.json`
+- `flywheel-metrics.json`
+
+## Readiness Checks
+
+Use these as heuristics, not hard rules:
+
+- 10-25 approved runs: useful first context card
+- 25-100 approved runs: first eval set and repeated failure modes
+- 100-300 approved runs: context compression and routing measurement
+- 500-1,500 high-quality examples: narrow adapter experiment candidate
+- 2,000-10,000+ examples: broader workflow-family corpus
+
+## What To Report
+
+When finishing, report:
+
+- traces exported
+- trainable examples exported
+- eval cases exported
+- context cards created
+- redaction pass rate
+- acceptance rate by workflow
+- workflows that should stay frontier-model/manual-review
+- workflows that may become SLM/adapter candidates later
+
+## Self-rewrite hook
+
+If users repeatedly ask for the same domain-specific fields, add them to a
+local context card or schema example instead of hard-coding them into this
+general skill.