AI assistants repeat the same mistakes because useful corrections disappear when the session ends. Adapt mines local Codex and Claude transcripts for repeated, durable guidance and promotes it — through hard safety gates — into a small, scoped, reversible preference layer that future agents actually recall.

_{Package & CLI id: adapt.}

It does not retrain the model, and it does not save private chain-of-thought. It learns things like "always run focused tests before reporting a broad build complete" — and refuses to learn things like "the service is down today."

The pipeline

flowchart LR
    T[local Codex + Claude<br/>transcripts] --> E[parse · canonicalize ·<br/>provenance filter]
    E --> S[deterministic extraction +<br/>LLM recall proposals]
    S --> A[authority checks<br/>origin quarantine]
    A --> M[immutable review manifest<br/>accepted / rejected / pending]
    M --> G[conformance gate] --> W[transactional<br/>Crypt apply]
    W --> R[scoped recall<br/>in future sessions]

Loading

Mining never writes rules directly. It emits a review manifest; only an adjudicated manifest can be applied; and apply is transactional with rollback.

What gets refused

Only authenticated user-origin evidence can create durable preference authority. Admission deterministically quarantines:

• assistant-authored narration and echoed tool/repository output (a prompt-injection lexical scan backs up the origin tags)
• permission or approval expansion, and anything that weakens security
• conflicts with the active AGENTS.md / CLAUDE.md / workspace rules, and contradictions with an active stored rule
• transient environment claims, forbidden scopes, unknown categories, duplicates, and rules too short to mean anything

Categories are a controlled taxonomy (workflow, verification, safety, architecture, tooling, code-style, documentation, model-routing); anything else is forced into review, never silently admitted.

Scoped, not global

Record types stop every lesson from becoming a global command:

Record type	Reach
standing_preference	Broad and durable — the only type eligible for the bounded always-on core
locked_decision	Binding, but only inside its declared scope
operational_playbook	Recalled when the task and scope match (the default)
episodic_fact	Supporting context, never a standing instruction
unclassified	Legacy/review state

Only root-scoped standing preferences compile into the always-on core; everything else stays recall-gated — so the preference layer never grows into another giant prompt.

Nothing applies unless it's exactly what was reviewed

Every manifest candidate carries its source session identities, per-transcript SHA-256 hashes, a payload SHA-256, rule type, scope, authority effect, and evidence links. Apply refuses: pending records, an edited payload whose hash no longer matches, a changed canonical rule pool, source sessions from another installation, out-of-manifest evidence, and authority-quarantined candidates.

And it's reversible: a run journal checkpoints every stage; safe resume reuses cached stages only while session identity still matches; apply captures snapshots first; rollback deletes only recorded IDs, restores snapshots, and runs PRAGMA integrity_check — no force flag bypasses a failed integrity proof.

Surfaces

Surface	Role	Status
Taste	durable preferences → Crypt	ships
Doctor	multiwriter conformance receipts (issue / validate)	ships; Blueprint/Forge checks not yet
Insights	failure/waste mining	deferred — not a product yet

Using it

python3 adapt.py --smoke                            # dry-run the whole pipeline
python3 adapt.py --incremental --manifest pending.json
python3 adapt.py --apply-from-manifest resolved.json
python3 adapt.py --compile-core path/to/core.json
python3 adapt.py --insights session-one.jsonl session-two.jsonl

python3 adapt.py doctor issue --out receipt.json
python3 adapt.py doctor validate --receipt receipt.json

Writes are opt-in (--apply); smoke & manifest generation stay dry-run. LLM proposal lanes are local (default) or minimax; every proposal is rebound to an exact canonical external-user event, then passes deterministic admission. --deterministic-only disables LLM recall explicitly. Tests: python3 -m pytest -q.

Repository layout

Path	Contents
src/adapt/	runtime package, policies & manifest schema
tests/	unit, contract & regression tests
eval/	offline evaluation and delivery-parity tooling
docs/	architecture, operations & historical plans

adapt.py is the intentionally small source-checkout entrypoint; implementation stays under src/adapt/.

Recent

• Every mined rule is now attributed to the machine that learned it; session IDs are installation-qualified so two machines can't collide.
• The Ollama lane was dropped entirely; the external lane is MiniMax at the proxy, with live e2e routing fixes.
• Lifecycle and verification fields land on the direct-rule path; scope dimensions and a planted-finding bench joined the eval harness.

Current limits

A standalone checkout depends on parent-workspace memory/session modules and an installed Crypt (workspace_runtime.py is the single import boundary; offline stubs exist but are barred from live applies). Model-assisted extraction needs a configured lane. Lexical contradiction detection catches direct polarity conflicts, not every semantic conflict. Doctor does not yet cover Blueprint or Forge.

---

_{Orthic Labs — local-first infrastructure for AI-assisted development.
Membrane · Cortex · Forge · Roundtable · Adapt · CutRight · claudecodeX}