import type { PromptDef } from "./index.js";

/**
 * V7 §5 — Cross-task L2 induction.
 *
 * Given a set of L1 traces that landed in the same signature bucket (similar
 * state + similar action), distill a candidate L2 policy that describes
 * "when you see X, prefer Y because Z". The candidate is still probationary
 * until the evaluator confirms it raises task success.
 *
 * Boundary contract (see `docs/GRANULARITY-AND-MEMORY-LAYERS.md` §6):
 * an L2 policy is **procedural** ("how to do it") — it MUST contain an
 * action template. Anything declarative ("the environment looks like X")
 * belongs to the L3 world model, not here. The system prompt explicitly
 * rejects environment-fact drift to keep the two layers semantically
 * orthogonal. Bumping the version to v2 captures that change.
 */
export const L2_INDUCTION_PROMPT: PromptDef = {
  id: "l2.induction",
  version: 2,
  description:
    "Distill an L2 policy (procedural sub-task strategy) from a cluster of similar L1 traces, with explicit boundaries against L3 world-model drift.",
  system: `You induce reusable **procedural policies** from agent experience.

A policy is a "how-to": "when you see condition X in the agent's state,
do action Y, verify with Z, watch out for caveat W." It is **NOT** a
description of the environment.

Input TRACES: a list of { state_summary, action, outcome, utility } records
that all share a similar state signature.

Produce ONE policy describing the action pattern. The policy must:
- Name a TRIGGER recognizable from the agent's STATE — a condition the
  agent can detect at the moment of decision (an error code, a missing
  file, a request shape). NOT a fact about the environment in general.
- Prescribe an ACTION template — a parameterized step or short step
  sequence. Templates over single exact commands. NOT a single example.
- Note at least one CAVEAT or failure mode observed in the traces — a
  step-level pitfall, NOT a generic environment taboo.
- Generalize across the input traces, not restate one of them.

──────────────────── Boundaries — what NOT to write ────────────────────

This output is a **procedural policy**, not an environment world model.
The world model lives in a separate layer (L3) generated by a different
prompt. Cross-contamination on either side dilutes both.

Do NOT write any of these — they belong to L3 (env world model), not here:
  - Topology facts: "Alpine containers ship musl libc"
                    "Python deps form a 3-layer stack"
                    "src/components/ holds React components"
  - Environment behavioural rules (in pure declarative form):
                    "binary wheels are incompatible with musl"
                    "the service reads config only at startup"
  - Environment taboos detached from a specific action choice:
                    "this directory is read-only"
                    "production tables shouldn't be DROPped lightly"

If a trace tells you the environment looks a certain way, FOLD that fact
INTO the trigger or caveat as a state-level CONDITION the agent can
check, not as a standalone description. Example:

  Wrong (drifts into env-fact):
    trigger:  "Alpine ships musl libc"
    caveats:  ["Python deps have a 3-layer stack"]

  Right (states it as actionable conditions):
    trigger:  "container is Alpine AND pip install fails with
               '<lib> not found' or 'header not found'"
    caveats:  ["if first apk add still fails, also check musl-vs-glibc
               wheel compatibility before retrying"]

──────────────────── Same fact, two framings ─────────────────────

If the underlying truth is "Alpine containers don't ship system dev
libs by default":

  Express here (procedural):
    "When pip install fails inside an Alpine container with a missing
     system library, run apk add <pkg>-dev then retry pip."

  Do NOT express here (declarative — that's L3's job):
    "Alpine container images ship only the pure-Python tier of the
     Python dependency stack."

──────────────────── Output ─────────────────────

Return JSON:
{
  "title": "short imperative title",
  "trigger": "state-level condition the agent can detect",
  "action": "templated step or step sequence",
  "rationale": "why this action works ON THESE TRACES (not why the
                environment behaves this way)",
  "caveats": ["step-level pitfall string", ...],
  "confidence": number in [0, 1],
  "support_trace_ids": ["tr_...", ...]
}`,
};