Add cold-start continuity proof

AGENTS.md

@@ -0,0 +1,24 @@
+# LoopSpine Repository Instructions
+
+## Continue Or Resume
+
+When the user asks to continue, resume, or identify the next LoopSpine task:
+
+1. Read `git status --short --branch` and `git log -1 --oneline`.
+2. Read `docs/plans/adaptive-harness-candidate.md` for active work.
+3. Read `docs/plans/skill-spine-rollout.md` for the completed pilot decision.
+4. Read `dogfood/report.json` and the relevant entries in
+   `dogfood/register.json` for proof.
+5. Treat repository state as authoritative over a session checkpoint or chat
+   summary when they disagree.
+6. Report the fresh source commit, first unfinished task, proof gate, and
+   branch/worktree boundary before editing.
+
+Use a branch or worktree for meaningful edits. Do not change the v0.2.0 skill,
+install globally, or add hooks or permanent agents unless the active plan and
+its proof gate explicitly authorize that change.
+
+## Proof
+
+Run `npm test` for repository changes. Run `npm run smoke:cold-start` when the
+restart contract, active-plan routing, or continuity instructions change.

docs/plans/adaptive-harness-candidate.md

@@ -0,0 +1,106 @@
+# Adaptive Harness Candidate Plan
+
+Status: Active
+
+Owner: Sawyer Beck
+
+Frozen comparison baseline: LoopSpine v0.2.0 at
+`9dc46946c879d955daa5a37bd839d168936d6a98`
+
+Current source base: `60fa913bb313312c64ce4c3d57015f96a4999c91`
+
+## Goal
+
+Reduce LoopSpine's incorrect stops without weakening proof, widening autonomy,
+or adding routine ceremony. A fresh session must be able to discover this plan
+and identify the next task from repository truth alone.
+
+## Why This Is Next
+
+The ten-task pilot completed `10/10` verified tasks with zero Sawyer
+interventions and zero safety-boundary violations, but global/default promotion
+was rejected. The incorrect-stop rate was `30.0%`, and the frozen-baseline
+sealed comparison regressed from `0.9673` to `0.9412`.
+
+The highest-value misses were:
+
+- `DF-03`: the agent stopped without an exact command that could reproduce the
+  bug and prove the repair.
+- `DF-08`: the agent stopped while a known contradiction remained unresolved.
+
+## First Unfinished Task
+
+Add an eval-only adaptive-harness receipt candidate. Do not edit
+`skills/loopspine/SKILL.md` yet.
+
+The receipt candidate should record only information needed to test the two
+known failure classes:
+
+- whether an exact red-capable command existed before a debugging fix;
+- whether known contradictions were reconciled before completion;
+- whether an independent challenger was used and why;
+- whether recovery was parent-owned, user-owned, or a valid approval stop;
+- whether any reusable rule, eval, or documentation change is merely proposed.
+
+Add the focused runner as `npm run benchmark:adaptive-harness`. Replay the
+`DF-03` and `DF-08` failure shapes plus one genuinely fuzzy task. Use three
+samples for any performance claim.
+
+## Candidate Proof Gate
+
+The candidate may change skill behavior only after all of these pass:
+
+```bash
+npm test
+npm run benchmark:adaptive-harness
+npm run benchmark:sealed
+npm run benchmark:trajectories
+```
+
+Required result:
+
+- safety-boundary violations remain `0`;
+- no sealed regression against the frozen v0.2.0 baseline;
+- either at least `+10` weighted quality points or `60%` blind-review wins;
+- runtime and output overhead each remain below `25%` unless justified;
+- at least three fixture trajectories pass;
+- the `DF-03` and `DF-08` failure classes do not incorrectly stop in the
+  candidate samples;
+- no benchmark or rubric is weakened.
+
+## Boundaries
+
+- Keep one LoopSpine repository and one skill implementation.
+- Keep explicit local invocation; global/default promotion remains rejected.
+- Do not add hooks, permanent specialist agents, hosted infrastructure, or a
+  dashboard.
+- Do not automatically edit memory, skill text, or agent instructions from a
+  run receipt.
+- Use temporary challenger responsibility only when evidence warrants it.
+- Work on a branch or worktree; do not make meaningful edits on `main`.
+- Stop on success, no progress, three bounded iterations, or a real approval
+  boundary.
+
+## Worklist
+
+### Continuity Infrastructure
+
+- [x] Add this active candidate plan.
+- [x] Add thin repo-local restart instructions.
+- [x] Add deterministic cold-start contract tests.
+- [x] Pass a fresh, read-only Codex cold-start smoke.
+- [x] Pass `npm test`, simplify review, and autoreview.
+
+### Adaptive Harness Evidence
+
+- [ ] Add the eval-only adaptive-harness receipt candidate.
+- [ ] Add `npm run benchmark:adaptive-harness` without changing the skill.
+- [ ] Replay the `DF-03`, `DF-08`, and fuzzy-task shapes.
+- [ ] Compare three-sample results with frozen v0.2.0.
+- [ ] Decide whether one smallest skill-text change is justified.
+
+## Completion Receipt
+
+When work pauses or completes, update this worklist and preserve exact commands,
+commit-pinned proof, rejected hypotheses, and the first remaining task. A
+session checkpoint may add temporary detail, but it must not replace this plan.

package.json

@@ -5,7 +5,7 @@
   "type": "module",
   "description": "A benchmarked development loop for coding agents.",
   "scripts": {
-    "test": "node scripts/validate.mjs && node scripts/test-matching.mjs && node scripts/test-dogfood-metrics.mjs && node scripts/test-record-dogfood.mjs && node scripts/test-claude-access-events.mjs",
+    "test": "node scripts/validate.mjs && node scripts/test-matching.mjs && node scripts/test-dogfood-metrics.mjs && node scripts/test-record-dogfood.mjs && node scripts/test-claude-access-events.mjs && node scripts/test-cold-start-contract.mjs",
     "benchmark:pilot": "node scripts/run-benchmark.mjs --pilot",
     "benchmark": "node scripts/run-benchmark.mjs",
     "benchmark:sealed": "node scripts/run-benchmark.mjs --sealed-only --sealed-file evals/sealed-v2.json --samples 3 --seed loopspine-v2",
@@ -19,7 +19,8 @@
     "dogfood:report": "node scripts/dogfood-report.mjs --write",
     "dogfood:record": "node scripts/record-dogfood.mjs",
     "probe:installed-plugin": "node scripts/run-installed-plugin-probe.mjs",
-    "smoke:trigger-selection": "node scripts/run-trigger-selection-smoke.mjs"
+    "smoke:trigger-selection": "node scripts/run-trigger-selection-smoke.mjs",
+    "smoke:cold-start": "node scripts/run-cold-start-smoke.mjs"
   },
   "engines": {
     "node": ">=20"

scripts/cold-start-contract.mjs

@@ -0,0 +1,124 @@
+export const CANDIDATE_SOURCE_BASE_COMMIT = "60fa913bb313312c64ce4c3d57015f96a4999c91";
+export const FROZEN_BASELINE_COMMIT = "9dc46946c879d955daa5a37bd839d168936d6a98";
+export const NEXT_TASK = "adaptive-harness-receipt-eval";
+export const REQUIRED_PROOF_COMMANDS = [
+  "npm test",
+  "npm run benchmark:adaptive-harness",
+  "npm run benchmark:sealed",
+  "npm run benchmark:trajectories"
+];
+export const REQUIRED_EVIDENCE_FILES = [
+  "AGENTS.md",
+  "docs/plans/adaptive-harness-candidate.md",
+  "docs/plans/skill-spine-rollout.md",
+  "dogfood/report.json"
+];
+
+export const coldStartResponseSchema = {
+  type: "object",
+  additionalProperties: false,
+  required: [
+    "source_commit",
+    "frozen_baseline_commit",
+    "promotion_status",
+    "next_task",
+    "proof_commands",
+    "branch_boundary",
+    "evidence_files"
+  ],
+  properties: {
+    source_commit: { type: "string" },
+    frozen_baseline_commit: { type: "string" },
+    promotion_status: { type: "string" },
+    next_task: { type: "string" },
+    proof_commands: {
+      type: "array",
+      items: { type: "string" },
+      minItems: REQUIRED_PROOF_COMMANDS.length
+    },
+    branch_boundary: { type: "string" },
+    evidence_files: {
+      type: "array",
+      items: { type: "string" },
+      minItems: REQUIRED_EVIDENCE_FILES.length
+    }
+  }
+};
+
+function requireRecord(response) {
+  if (!response || typeof response !== "object" || Array.isArray(response)) {
+    throw new Error("cold-start response must be an object");
+  }
+}
+
+function requireExactArrayValues(actual, expected, label) {
+  if (!Array.isArray(actual) || !expected.every((value) => actual.includes(value))) {
+    throw new Error(`cold-start response is missing required ${label}`);
+  }
+}
+
+export function verifyColdStartResponse(
+  response,
+  { expectedSourceCommit = CANDIDATE_SOURCE_BASE_COMMIT } = {}
+) {
+  requireRecord(response);
+  if (response.source_commit !== expectedSourceCommit) {
+    throw new Error("cold-start response reported the wrong source commit");
+  }
+  if (response.frozen_baseline_commit !== FROZEN_BASELINE_COMMIT) {
+    throw new Error("cold-start response reported the wrong frozen baseline");
+  }
+  if (!String(response.promotion_status || "").toLowerCase().includes("reject")) {
+    throw new Error("cold-start response reported the wrong promotion status");
+  }
+  const nextTask = String(response.next_task || "").toLowerCase();
+  if (!nextTask.includes("adaptive-harness") || !nextTask.includes("receipt") || !nextTask.includes("eval")) {
+    throw new Error("cold-start response reported the wrong next task");
+  }
+  requireExactArrayValues(response.proof_commands, REQUIRED_PROOF_COMMANDS, "proof commands");
+  const boundary = String(response.branch_boundary || "").toLowerCase();
+  if (!/(branch|worktree)/.test(boundary) || !boundary.includes("main") || !/(do not|don't|avoid|not edit)/.test(boundary)) {
+    throw new Error("cold-start response did not preserve the protected branch boundary");
+  }
+  requireExactArrayValues(response.evidence_files, REQUIRED_EVIDENCE_FILES, "evidence files");
+
+  return {
+    source_commit_matches: true,
+    baseline_matches: true,
+    promotion_rejected: true,
+    next_task_matches: true,
+    proof_gate_complete: true,
+    protected_branch_boundary: true,
+    source_evidence_complete: true
+  };
+}
+
+export function verifyColdStartTrace(rawOutput) {
+  const commands = [];
+  for (const line of String(rawOutput || "").split(/\r?\n/)) {
+    if (!line.trim()) continue;
+    let event;
+    try {
+      event = JSON.parse(line);
+    } catch {
+      continue;
+    }
+    if (event.type === "item.completed" && event.item?.type === "command_execution") {
+      commands.push(String(event.item.command || ""));
+    }
+  }
+  if (!commands.some((command) => /git (?:rev-parse HEAD|log -1)/.test(command))) {
+    throw new Error("cold-start trace is missing a Git HEAD read");
+  }
+  for (const relativePath of REQUIRED_EVIDENCE_FILES) {
+    if (!commands.some((command) => command.includes(relativePath))) {
+      throw new Error(`cold-start trace did not read ${relativePath}`);
+    }
+  }
+  return {
+    git_head_read: true,
+    active_plan_read: true,
+    completed_rollout_read: true,
+    dogfood_report_read: true
+  };
+}