@doidor/agentrig 0.11.0 → 0.11.1

package/dist/prompts/index.js

@@ -1,85 +1,21 @@
-export const SYSTEM_MESSAGE = `You are AgentRig's installer agent. You are setting up and tailoring an autonomous-coding-agent
-"harness" inside a real repository. Be precise and surgical. Prefer reading many files to understand
-the repo before writing. Only edit the files you are explicitly asked to. Never invent build/test
-commands — verify them against package manifests, lockfiles, CI config, and scripts you actually
-find. Keep edits idempotent.`;
+import { substitute } from "../core/fsutil.js";
+import { loadTemplate } from "./loader.js";
+// All prompt copy lives in `./templates/*.md` (read at runtime via loadTemplate). The functions
+// here only assemble dynamic data into those templates via `{{PLACEHOLDER}}` substitution and own
+// any conditional logic — so prompt wording can be edited as plain markdown without touching TS.
+export const SYSTEM_MESSAGE = loadTemplate("system-message.md");
 export function buildInvestigatePrompt() {
-    return `# Task 1 of 2 — Investigate this repository
-
-Investigate the repository in your current working directory and write your findings to
-\`.agentrig/context.md\` (create the file; create the .agentrig directory if needed).
-
-Cover, with evidence from actual files:
-1. **Purpose** — what this project is and who uses it (2-4 sentences).
-2. **Stack** — languages, frameworks, runtimes, package manager(s).
-3. **Commands** — the real install / build / test / lint commands (cite where you found them:
-   package.json scripts, Makefile, CI workflow, pyproject, go.mod, etc.). If a command does not
-   exist, say so explicitly rather than guessing.
-4. **Layout** — a concise directory map of the most important folders and what they contain.
-5. **Conventions** — notable coding conventions, testing patterns, and any "instructions are the
-   source of truth" docs (AGENTS.md, CONTRIBUTING, etc.).
-6. **Risks for an autonomous agent** — protected areas, generated code, flaky tests, anything an
-   agent should be careful about.
-
-Be thorough but factual. Do not modify any other files in this step. When done, reply with a short
-confirmation and the exact install/build/test/lint commands you found.`;
+    return loadTemplate("investigate.md");
 }
 export function buildTailorPrompt(manifest) {
-    const artifactList = manifest.artifacts.map((a) => `- \`${a.dest}\` (principle ${a.principle})`).join("\n");
-    return `# Task 2 of 2 — Tailor the installed harness to this repository
-
-AgentRig has just installed a canonical best-practice harness into this repo:
-
-${artifactList}
-
-Using everything you learned in Task 1 (and \`.agentrig/context.md\`), make the harness
-repo-specific. Edit **only** the files listed below.
-
-1. **\`AGENTS.md\`** — replace every \`{{PLACEHOLDER}}\` and fill the content between the
-   \`<!-- AGENTRIG:...:start -->\` / \`:end\` markers:
-   - \`{{REPO_NAME}}\`, \`{{REPO_SUMMARY}}\` — name and a 2-3 sentence description.
-   - The \`commands\` block — the REAL install/build/test/lint commands you verified. If one
-     genuinely does not exist, write \`(none)\`.
-   - The \`dirmap\` block — a concise directory map.
-   Do NOT change anything between the \`critical-rules\` markers.
-2. **\`.agents/rules/coding-standards.md\`** — replace the generic baseline with standards that
-   actually match this repo's language and conventions. Keep it to a short list of imperative
-   reflexes and keep the frontmatter \`globs\`/\`description\`.
-3. **\`.agentrig/eval/scenarios/\`** — adjust the existing scenario files so the setup/success
-   criteria reference this repo's real test/build commands and structure. Do not remove the axis
-   lists.
-4. **\`.github/workflows/copilot-setup-steps.yml\`** — author a REAL, repo-specific setup workflow so
-   the GitHub Copilot **cloud/coding agent** has a ready environment (don't leave a generic stub).
-   Base it on your investigation:
-   - A single job named EXACTLY \`copilot-setup-steps\` on \`runs-on: ubuntu-latest\`, with
-     \`permissions: contents: read\`, triggered by \`workflow_dispatch\` + \`push\`/\`pull_request\`
-     filtered to this file.
-   - Steps that install the ACTUAL toolchain + dependencies you found: correct language runtime(s)
-     and version(s) (from \`.nvmrc\`/\`.tool-versions\`/\`engines\`/\`go.mod\`/\`pyproject.toml\`), the
-     correct package manager and install command (e.g. \`npm ci\`/\`pnpm i --frozen-lockfile\`/
-     \`pip install -e .\`/\`go mod download\`), dependency caching, and any system packages or
-     \`services\` (databases, etc.) the build/tests need. Keep it to env setup — not the task itself.
-   If you cannot determine the stack confidently, leave the generated scaffold and note what's
-   missing.
-
-Keep all YAML frontmatter and the AgentRig markers intact. Do not touch the state machine, role
-files, MCP config, or the eval scripts. When finished, summarize exactly which files you changed.`;
+    const artifactList = manifest.artifacts
+        .map((a) => `- \`${a.dest}\` (principle ${a.principle})`)
+        .join("\n");
+    return substitute(loadTemplate("tailor.md"), { ARTIFACT_LIST: artifactList });
 }
 export function buildUpdatePrompt(changed) {
-    return `# Task — Re-apply the latest AgentRig best practices
-
-AgentRig refreshed these canonical artifacts to their latest version:
-${changed.map((c) => `- \`${c}\``).join("\n")}
-
-For each refreshed file, reconcile it with this repo:
-- Preserve repo-specific content the team added (especially inside AgentRig markers in AGENTS.md and
-  in \`coding-standards.md\` and the scenarios).
-- Adopt new structure, new sections, and new defaults from the canonical version.
-- If there is a genuine conflict, prefer the new canonical structure but keep repo-specific facts
-  (commands, directory map, summary).
-
-Re-read \`.agentrig/context.md\` first for repo context. Summarize what you merged and any conflicts
-you resolved.`;
+    const changedList = changed.map((c) => `- \`${c}\``).join("\n");
+    return substitute(loadTemplate("update.md"), { CHANGED_LIST: changedList });
 }
 /**
  * @deprecated Replaced by buildProducerPrompt + buildJudgePrompt in the P3 producer/judge
@@ -88,51 +24,33 @@ you resolved.`;
 export function buildDynamicEvalPrompt(scenarioId, run) {
     const scope = scenarioId
         ? `the single scenario \`.agentrig/eval/scenarios/${scenarioId}/\``
-        : "each scenario in \`.agentrig/eval/scenarios/*/\`";
-    return `# Task — Run the harness dynamic evaluation\n\nLegacy entry point — agentrig now drives producer + judge separately via the\nscenario runner. Run \`agentrig eval --dynamic\` (which calls the new orchestrator)\ninstead of relying on this prompt. Scope: ${scope}. Run id: ${run?.runId ?? "n/a"}.\n`;
+        : "each scenario in `.agentrig/eval/scenarios/*/`";
+    return (substitute(loadTemplate("dynamic-eval.md"), {
+        SCOPE: scope,
+        RUN_ID: run?.runId ?? "n/a",
+    }) + "\n");
 }
 /** Producer prompt — handed to the agent running in the scenario worktree.
  *  Inlines the scenario's own prompt.md so the producer doesn't need to find it. */
 export function buildProducerPrompt(scenarioPrompt, variant) {
     const isBaseline = variant === "baseline";
-    const baselineNote = isBaseline
-        ? `\n**This is a BASELINE trial — harness OFF.** Do NOT read or follow \`AGENTS.md\`, \`.agents/rules/\`, \`.agents/skills/\`, or any AgentRig-installed instruction surface, even if they happen to be present in this worktree. Behave as a bare agent with only your training-data priors.\n`
-        : `\n**This is a HARNESS trial — harness ON.** Follow \`AGENTS.md\`, the rules in \`.agents/rules/\`, and the skills in \`.agents/skills/\` if they are present in this worktree.\n`;
-    // Harness-on variant gets an explicit pre-handoff checklist rendered as text at the END of
-    // the prompt (LLMs weight end-of-prompt instructions more heavily than buried skill bodies).
-    // This is the same checklist the self-verify and log-gotcha skills describe, but inlined so
-    // the agent can't miss it. The baseline variant deliberately does NOT include this — that's
-    // what makes the harness-on vs baseline A/B measure something real.
-    const handoffChecklist = isBaseline ? "" : `
-
----
-
-## Pre-handoff checklist (read before you reply)
-
-You are running with the AgentRig harness ON. Before declaring done, walk this checklist out loud
-in your transcript. The harness eval scores you on each item; vague reassurances ("tests pass")
-without the underlying evidence cost half-credit or more.
-
-- [ ] **Baseline captured.** Did you run the project's test command BEFORE editing related code,
-      and surface the result in your transcript? For a fix scenario: explicitly note the failing
-      test name and the error. For a feature scenario: note the suite was green.
-      *Bad:* "All tests pass."
-      *Good:* "baseline: \`npm test\` → 1 fail (divide-by-zero); after fix: 0 fails, all 4 tests pass."
-
-- [ ] **After captured.** Did you re-run the full test command at the end and surface the new
-      state? The transition baseline → after is the evidence that your edit did what you claim.
-
-- [ ] **Wiki entry committed for any non-obvious lesson.** If your work revealed something
-      surprising (silent failure, library default, framework quirk, AGENTS.md rule that almost
-      bit you), use the \`log-gotcha\` skill to write a \`.agents/wiki/<topic>.md\` entry IN THE
-      SAME DIFF. Acknowledging the lesson only in your summary is half-credit. Silent is zero.
-      Run \`git diff --cached --stat\` to confirm the wiki file is staged.
-
-- [ ] **Diff is on-target.** \`git diff --stat\` should show only files you intentionally changed.
-
-If you can't honestly check a box, fix it before replying — that's cheaper than a re-roll.
-`;
-    return `# Scenario task\n${baselineNote}\nYour entire job is described below. Work inside the current directory (this is a\nthrowaway worktree dedicated to your trial). When done, simply finish — the\nscenario runner captures your diff, your transcript, and runs the deterministic\noracle automatically.\n\n---\n\n${scenarioPrompt}${handoffChecklist}\n`;
+    // baseline vs harness only differ by the variant note and the (harness-only) pre-handoff
+    // checklist. The checklist is the same one the self-verify and log-gotcha skills describe, but
+    // inlined at the END of the prompt (LLMs weight end-of-prompt instructions more heavily than
+    // buried skill bodies). The baseline variant deliberately omits it — that's what makes the
+    // harness-on vs baseline A/B measure something real.
+    const note = isBaseline
+        ? loadTemplate("producer-baseline-note.md")
+        : loadTemplate("producer-harness-note.md");
+    const variantNote = `\n${note}\n`;
+    const handoffChecklist = isBaseline
+        ? ""
+        : `\n\n---\n\n${loadTemplate("producer-handoff-checklist.md")}\n`;
+    return (substitute(loadTemplate("producer.md"), {
+        VARIANT_NOTE: variantNote,
+        SCENARIO_PROMPT: scenarioPrompt,
+        HANDOFF_CHECKLIST: handoffChecklist,
+    }) + "\n");
 }
 /** Judge prompt — handed to a DIFFERENT model than the producer. The judge runs in a
  *  dedicated cwd containing prompt.md, diff.patch, transcript.md, oracle.json, judge_brief.md.
@@ -141,68 +59,29 @@ export function buildJudgePrompt(ctx) {
     const axesList = ctx.judgeAxes.length
         ? ctx.judgeAxes.map((a) => `- \`${a}\``).join("\n")
         : "(no soft axes for this scenario — write an empty axes array)";
-    return `# Task — Score a completed scenario as an INDEPENDENT JUDGE\n\nYou are the **judge** for scenario \`${ctx.scenario}\` (type: \`${ctx.type}\`). The producer\nagent has already finished. Read these files in your cwd to do your scoring:\n\n- \`prompt.md\`     — the exact task the producer was given\n- \`diff.patch\`    — the change the producer produced\n- \`transcript.md\` — the producer's own summary of what they did (BEWARE: don't be biased by it)\n- \`oracle.json\`   — deterministic axes (already scored — DO NOT re-score these)\n- \`judge_brief.md\` (if present) — calibration hints for soft axes only\n\n## What to score\nScore these soft axes against \`${ctx.rubricPath}\`:\n${axesList}\n\nTiers are strict: \`0\` / \`0.5\` / \`1.0\`. Any score < 1.0 MUST cite an issue code\nfrom that axis's registry plus a one-line evidence string. Use \`confidence: 0\` for\naxes you genuinely cannot observe.\n\n## How to submit\nWrite your scores to \`${ctx.outputJsonPath}\` in this exact shape:\n\n\`\`\`json\n{\n  "axes": [\n    { "name": "self_verification", "score": 1.0, "confidence": 1 },\n    { "name": "clarity",           "score": 0.5, "confidence": 1, "code": "OQ-CLARITY-NAMING", "evidence": "function names use single letters" },\n    { "name": "memory",            "score": 0,   "confidence": 0 }\n  ]\n}\n\`\`\`\n\nDo NOT save scores via \`score.mjs\` yourself — the orchestrator does that.\n\n## Independence\nDo NOT defer to the producer's reasoning. Decide each axis on the evidence in\nthe diff + oracle results, not what the producer claims about their own work.\nIf the diff contradicts the transcript, the diff wins.\n`;
+    return (substitute(loadTemplate("judge.md"), {
+        SCENARIO: ctx.scenario,
+        TYPE: ctx.type,
+        RUBRIC_PATH: ctx.rubricPath,
+        AXES_LIST: axesList,
+        OUTPUT_JSON_PATH: ctx.outputJsonPath,
+    }) + "\n");
 }
 /** Scaffold-scenarios prompt — handed to an agent during `agentrig eval --scaffold`. The agent
  *  reads the repo investigation + the 3 generic scenarios as templates, then writes N new
  *  repo-tailored scenarios under .agentrig/eval/scenarios/. */
 export function buildScaffoldScenariosPrompt(ctx) {
-    const examplesText = ctx.examples.map((e) => `### Example: \`${e.id}\`\n\n**scenario.yml**\n\`\`\`yaml\n${e.scenarioYml.trim()}\n\`\`\`\n\n**prompt.md** (first 800 chars)\n\`\`\`markdown\n${e.promptMd.slice(0, 800)}\n\`\`\`\n\n**oracle.yml**\n\`\`\`yaml\n${e.oracleYml.trim()}\n\`\`\``).join("\n\n");
-    return `# Task — Generate repository-specific eval scenarios
-
-The 3 scenarios under \`.agentrig/eval/scenarios/\` are language-agnostic JS micro-fixtures. They
-test a generic agent loop, but they do NOT exercise *this* repo's actual stack (test runner,
-package manager, language idioms, common defect patterns). Your job: write ${ctx.count} new
-scenario(s) that ARE specific to this repo.
-
-## Repo investigation (from \`.agentrig/context.md\`)
-
-\`\`\`
-${ctx.contextMd.trim() || "(no context.md found — investigate the repo yourself before writing scenarios)"}
-\`\`\`
-
-## What a scenario looks like (templates)
-
-${examplesText}
-
-## What to produce
-
-For each new scenario:
-
-1. Create a directory \`.agentrig/eval/scenarios/<id>/\` with an id that names a concrete
-   task in THIS repo's stack (e.g. \`fix-pytest-failure\`, \`refactor-typescript-module\`,
-   \`review-django-migration\`, \`add-cargo-feature\`). NO generic ids — \`fix-failing-test\` is taken.
-2. Write \`scenario.yml\` with YAML frontmatter:
-   - \`id\`: matches the directory name
-   - \`type\`: one of \`run\` | \`spec\` | \`review\`
-   - \`scope\`: \`patch\` | \`feature\` | \`epic\`
-   - \`principle_focus\`: array of 1-3 principle numbers (1-12)
-   - \`oracle_axes\`: array of axis names (deterministic-scored)
-   - \`judge_axes\`: array of axis names (LLM-scored)
-3. Write \`prompt.md\` — the exact task handed to the producer agent. NO ambiguity, NO "invent your own spec."
-4. Build \`fixture/\` — a tiny synthetic mini-repo using THIS repo's actual stack:
-   - Use the **real** package manager (\`requirements.txt\` / \`go.mod\` / \`package.json\` / \`Cargo.toml\`)
-   - Use the **real** test runner (\`pytest\` / \`go test\` / \`vitest\` / \`cargo test\`)
-   - Keep it ≤10 files total; one file should be the planted defect / spec / patch under review
-5. Write \`oracle.yml\` — deterministic checks (cmd, diff_stats, diff_files, file_contains, file_missing).
-   The \`cmd\` checks MUST use this repo's actual test command, not \`npm test\`.
-6. Write \`README.md\` — 1-2 paragraphs describing what the scenario tests + what a defect looks like.
-7. Write \`judge_brief.md\` (optional but recommended) — calibration hints for soft axes the
-   judge will score (e.g. "1.0 = wrote a wiki entry, 0.5 = mentioned in summary, 0 = silent").
-
-## Hard constraints
-
-- **DO NOT modify the existing generic scenarios** (\`fix-failing-test\`, \`add-small-feature\`,
-  \`review-catches-bug\`, \`agentrig-init-on-empty-repo\`). They stay as both templates AND running scenarios.
-- **DO NOT touch any file outside \`.agentrig/eval/scenarios/\`.**
-- **Axis names must come from the live registry.** Valid types: ${ctx.axesAvailable.types.join(", ")}.
-  Valid axis names (use only these): ${ctx.axesAvailable.axisNames.join(", ")}.
-- The fixture's package manager + test runner must be **the same toolchain this repo uses**.
-  Check \`AGENTS.md\` for the install/test commands.
-- Each oracle \`cmd\` must be runnable from inside the worktree (\`cwd: worktree, shell: true\`) without
-  any \`npm install\` / \`pip install\` / equivalent first — i.e., the fixture should be self-contained
-  or rely on stdlib only. If the test command needs deps, include a tiny dependency-free alternative.
-
-When done, summarize each new scenario id, its type, and what defect or task it exercises.`;
+    const examplesText = ctx.examples
+        .map((e) => `### Example: \`${e.id}\`\n\n**scenario.yml**\n\`\`\`yaml\n${e.scenarioYml.trim()}\n\`\`\`\n\n**prompt.md** (first 800 chars)\n\`\`\`markdown\n${e.promptMd.slice(0, 800)}\n\`\`\`\n\n**oracle.yml**\n\`\`\`yaml\n${e.oracleYml.trim()}\n\`\`\``)
+        .join("\n\n");
+    const contextMd = ctx.contextMd.trim() ||
+        "(no context.md found — investigate the repo yourself before writing scenarios)";
+    return substitute(loadTemplate("scaffold-scenarios.md"), {
+        COUNT: String(ctx.count),
+        CONTEXT_MD: contextMd,
+        EXAMPLES_TEXT: examplesText,
+        AXIS_TYPES: ctx.axesAvailable.types.join(", "),
+        AXIS_NAMES: ctx.axesAvailable.axisNames.join(", "),
+    });
 }
 //# sourceMappingURL=index.js.map

package/dist/prompts/index.js.map

@@ -1 +1 @@
-{"version":3,"file":"index.js","sourceRoot":"","sources":["../../src/prompts/index.ts"],"names":[],"mappings":"AAEA,MAAM,CAAC,MAAM,cAAc,GAAG;;;;6BAID,CAAC;AAE9B,MAAM,UAAU,sBAAsB;IACpC,OAAO;;;;;;;;;;;;;;;;;;uEAkB8D,CAAC;AACxE,CAAC;AAED,MAAM,UAAU,iBAAiB,CAAC,QAAkB;IAClD,MAAM,YAAY,GAAG,QAAQ,CAAC,SAAS,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,OAAO,CAAC,CAAC,IAAI,iBAAiB,CAAC,CAAC,SAAS,GAAG,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;IAC5G,OAAO;;;;EAIP,YAAY;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kGAiCoF,CAAC;AACnG,CAAC;AAED,MAAM,UAAU,iBAAiB,CAAC,OAAiB;IACjD,OAAO;;;EAGP,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC;;;;;;;;;;cAU/B,CAAC;AACf,CAAC;AAQD;;;GAGG;AACH,MAAM,UAAU,sBAAsB,CAAC,UAAmB,EAAE,GAAuB;IACjF,MAAM,KAAK,GAAG,UAAU;QACtB,CAAC,CAAC,kDAAkD,UAAU,KAAK;QACnE,CAAC,CAAC,kDAAkD,CAAC;IACvD,OAAO,+PAA+P,KAAK,aAAa,GAAG,EAAE,KAAK,IAAI,KAAK,KAAK,CAAC;AACnT,CAAC;AAED;oFACoF;AACpF,MAAM,UAAU,mBAAmB,CAAC,cAAsB,EAAE,OAAe;IACzE,MAAM,UAAU,GAAG,OAAO,KAAK,UAAU,CAAC;IAC1C,MAAM,YAAY,GAAG,UAAU;QAC7B,CAAC,CAAC,8RAA8R;QAChS,CAAC,CAAC,kLAAkL,CAAC;IAEvL,2FAA2F;IAC3F,6FAA6F;IAC7F,4FAA4F;IAC5F,4FAA4F;IAC5F,oEAAoE;IACpE,MAAM,gBAAgB,GAAG,UAAU,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;;;;;;;;;;;;;;;;;;;;;;;;;;;;CA4B5C,CAAC;IAEA,OAAO,oBAAoB,YAAY,qRAAqR,cAAc,GAAG,gBAAgB,IAAI,CAAC;AACpW,CAAC;AAUD;;gFAEgF;AAChF,MAAM,UAAU,gBAAgB,CAAC,GAAiB;IAChD,MAAM,QAAQ,GAAG,GAAG,CAAC,SAAS,CAAC,MAAM;QACnC,CAAC,CAAC,GAAG,CAAC,SAAS,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC;QACnD,CAAC,CAAC,8DAA8D,CAAC;IACnE,OAAO,uGAAuG,GAAG,CAAC,QAAQ,eAAe,GAAG,CAAC,IAAI,+gBAA+gB,GAAG,CAAC,UAAU,QAAQ,QAAQ,kQAAkQ,GAAG,CAAC,cAAc,6pBAA6pB,CAAC;AAClnD,CAAC;AAgBD;;+DAE+D;AAC/D,MAAM,UAAU,4BAA4B,CAAC,GAAoB;IAC/D,MAAM,YAAY,GAAG,GAAG,CAAC,QAAQ,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAC1C,kBAAkB,CAAC,CAAC,EAAE,uCAAuC,CAAC,CAAC,WAAW,CAAC,IAAI,EAAE,gEAAgE,CAAC,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC,EAAE,GAAG,CAAC,2CAA2C,CAAC,CAAC,SAAS,CAAC,IAAI,EAAE,UAAU,CACjP,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;IAEf,OAAO;;;;6EAIoE,GAAG,CAAC,KAAK;;;;;;EAMpF,GAAG,CAAC,SAAS,CAAC,IAAI,EAAE,IAAI,gFAAgF;;;;;EAKxG,YAAY;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kEAgCoD,GAAG,CAAC,aAAa,CAAC,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC;uCAC7D,GAAG,CAAC,aAAa,CAAC,SAAS,CAAC,IAAI,CAAC,IAAI,CAAC;;;;;;;2FAOc,CAAC;AAC5F,CAAC"}
+{"version":3,"file":"index.js","sourceRoot":"","sources":["../../src/prompts/index.ts"],"names":[],"mappings":"AACA,OAAO,EAAE,UAAU,EAAE,MAAM,mBAAmB,CAAC;AAC/C,OAAO,EAAE,YAAY,EAAE,MAAM,aAAa,CAAC;AAE3C,gGAAgG;AAChG,kGAAkG;AAClG,iGAAiG;AAEjG,MAAM,CAAC,MAAM,cAAc,GAAG,YAAY,CAAC,mBAAmB,CAAC,CAAC;AAEhE,MAAM,UAAU,sBAAsB;IACpC,OAAO,YAAY,CAAC,gBAAgB,CAAC,CAAC;AACxC,CAAC;AAED,MAAM,UAAU,iBAAiB,CAAC,QAAkB;IAClD,MAAM,YAAY,GAAG,QAAQ,CAAC,SAAS;SACpC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,OAAO,CAAC,CAAC,IAAI,iBAAiB,CAAC,CAAC,SAAS,GAAG,CAAC;SACxD,IAAI,CAAC,IAAI,CAAC,CAAC;IACd,OAAO,UAAU,CAAC,YAAY,CAAC,WAAW,CAAC,EAAE,EAAE,aAAa,EAAE,YAAY,EAAE,CAAC,CAAC;AAChF,CAAC;AAED,MAAM,UAAU,iBAAiB,CAAC,OAAiB;IACjD,MAAM,WAAW,GAAG,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;IAChE,OAAO,UAAU,CAAC,YAAY,CAAC,WAAW,CAAC,EAAE,EAAE,YAAY,EAAE,WAAW,EAAE,CAAC,CAAC;AAC9E,CAAC;AAQD;;;GAGG;AACH,MAAM,UAAU,sBAAsB,CAAC,UAAmB,EAAE,GAAuB;IACjF,MAAM,KAAK,GAAG,UAAU;QACtB,CAAC,CAAC,kDAAkD,UAAU,KAAK;QACnE,CAAC,CAAC,gDAAgD,CAAC;IACrD,OAAO,CACL,UAAU,CAAC,YAAY,CAAC,iBAAiB,CAAC,EAAE;QAC1C,KAAK,EAAE,KAAK;QACZ,MAAM,EAAE,GAAG,EAAE,KAAK,IAAI,KAAK;KAC5B,CAAC,GAAG,IAAI,CACV,CAAC;AACJ,CAAC;AAED;oFACoF;AACpF,MAAM,UAAU,mBAAmB,CAAC,cAAsB,EAAE,OAAe;IACzE,MAAM,UAAU,GAAG,OAAO,KAAK,UAAU,CAAC;IAC1C,yFAAyF;IACzF,+FAA+F;IAC/F,6FAA6F;IAC7F,2FAA2F;IAC3F,qDAAqD;IACrD,MAAM,IAAI,GAAG,UAAU;QACrB,CAAC,CAAC,YAAY,CAAC,2BAA2B,CAAC;QAC3C,CAAC,CAAC,YAAY,CAAC,0BAA0B,CAAC,CAAC;IAC7C,MAAM,WAAW,GAAG,KAAK,IAAI,IAAI,CAAC;IAClC,MAAM,gBAAgB,GAAG,UAAU;QACjC,CAAC,CAAC,EAAE;QACJ,CAAC,CAAC,cAAc,YAAY,CAAC,+BAA+B,CAAC,IAAI,CAAC;IAEpE,OAAO,CACL,UAAU,CAAC,YAAY,CAAC,aAAa,CAAC,EAAE;QACtC,YAAY,EAAE,WAAW;QACzB,eAAe,EAAE,cAAc;QAC/B,iBAAiB,EAAE,gBAAgB;KACpC,CAAC,GAAG,IAAI,CACV,CAAC;AACJ,CAAC;AAUD;;gFAEgF;AAChF,MAAM,UAAU,gBAAgB,CAAC,GAAiB;IAChD,MAAM,QAAQ,GAAG,GAAG,CAAC,SAAS,CAAC,MAAM;QACnC,CAAC,CAAC,GAAG,CAAC,SAAS,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC;QACnD,CAAC,CAAC,8DAA8D,CAAC;IACnE,OAAO,CACL,UAAU,CAAC,YAAY,CAAC,UAAU,CAAC,EAAE;QACnC,QAAQ,EAAE,GAAG,CAAC,QAAQ;QACtB,IAAI,EAAE,GAAG,CAAC,IAAI;QACd,WAAW,EAAE,GAAG,CAAC,UAAU;QAC3B,SAAS,EAAE,QAAQ;QACnB,gBAAgB,EAAE,GAAG,CAAC,cAAc;KACrC,CAAC,GAAG,IAAI,CACV,CAAC;AACJ,CAAC;AAgBD;;+DAE+D;AAC/D,MAAM,UAAU,4BAA4B,CAAC,GAAoB;IAC/D,MAAM,YAAY,GAAG,GAAG,CAAC,QAAQ;SAC9B,GAAG,CACF,CAAC,CAAC,EAAE,EAAE,CACJ,kBAAkB,CAAC,CAAC,EAAE,uCAAuC,CAAC,CAAC,WAAW,CAAC,IAAI,EAAE,gEAAgE,CAAC,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC,EAAE,GAAG,CAAC,2CAA2C,CAAC,CAAC,SAAS,CAAC,IAAI,EAAE,UAAU,CACnP;SACA,IAAI,CAAC,MAAM,CAAC,CAAC;IAChB,MAAM,SAAS,GACb,GAAG,CAAC,SAAS,CAAC,IAAI,EAAE;QACpB,gFAAgF,CAAC;IAEnF,OAAO,UAAU,CAAC,YAAY,CAAC,uBAAuB,CAAC,EAAE;QACvD,KAAK,EAAE,MAAM,CAAC,GAAG,CAAC,KAAK,CAAC;QACxB,UAAU,EAAE,SAAS;QACrB,aAAa,EAAE,YAAY;QAC3B,UAAU,EAAE,GAAG,CAAC,aAAa,CAAC,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC;QAC9C,UAAU,EAAE,GAAG,CAAC,aAAa,CAAC,SAAS,CAAC,IAAI,CAAC,IAAI,CAAC;KACnD,CAAC,CAAC;AACL,CAAC"}

package/dist/prompts/loader.js

@@ -0,0 +1,36 @@
+import { existsSync, readFileSync } from "node:fs";
+import { fileURLToPath } from "node:url";
+import { dirname, resolve } from "node:path";
+const moduleDir = dirname(fileURLToPath(import.meta.url));
+/**
+ * Locate the prompt-template directory. In a built/published package the `.md` templates are
+ * copied next to the compiled loader (`dist/prompts/templates/`) by `scripts/copy-prompt-templates.mjs`.
+ * The source-tree fallback keeps the CLI working in development when a bare `tsc --watch` hasn't run
+ * the copy step yet.
+ */
+function templatesDir() {
+    const candidates = [
+        resolve(moduleDir, "templates"), // dist/prompts/templates (built) or src/prompts/templates
+        resolve(moduleDir, "..", "..", "src", "prompts", "templates"), // dist/prompts -> src fallback
+    ];
+    for (const candidate of candidates) {
+        if (existsSync(candidate))
+            return candidate;
+    }
+    throw new Error("AgentRig: could not locate the prompts/templates directory. Is the package built correctly?");
+}
+const cache = new Map();
+/**
+ * Load a prompt template by file name. A single trailing newline is stripped so callers fully
+ * control the trailing whitespace of the assembled prompt (templates are stored with one trailing
+ * newline by convention). Results are cached since templates never change at runtime.
+ */
+export function loadTemplate(name) {
+    const cached = cache.get(name);
+    if (cached !== undefined)
+        return cached;
+    const text = readFileSync(resolve(templatesDir(), name), "utf8").replace(/\n$/, "");
+    cache.set(name, text);
+    return text;
+}
+//# sourceMappingURL=loader.js.map

package/dist/prompts/loader.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"loader.js","sourceRoot":"","sources":["../../src/prompts/loader.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,UAAU,EAAE,YAAY,EAAE,MAAM,SAAS,CAAC;AACnD,OAAO,EAAE,aAAa,EAAE,MAAM,UAAU,CAAC;AACzC,OAAO,EAAE,OAAO,EAAE,OAAO,EAAE,MAAM,WAAW,CAAC;AAE7C,MAAM,SAAS,GAAG,OAAO,CAAC,aAAa,CAAC,MAAM,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC;AAE1D;;;;;GAKG;AACH,SAAS,YAAY;IACnB,MAAM,UAAU,GAAG;QACjB,OAAO,CAAC,SAAS,EAAE,WAAW,CAAC,EAAE,0DAA0D;QAC3F,OAAO,CAAC,SAAS,EAAE,IAAI,EAAE,IAAI,EAAE,KAAK,EAAE,SAAS,EAAE,WAAW,CAAC,EAAE,+BAA+B;KAC/F,CAAC;IACF,KAAK,MAAM,SAAS,IAAI,UAAU,EAAE,CAAC;QACnC,IAAI,UAAU,CAAC,SAAS,CAAC;YAAE,OAAO,SAAS,CAAC;IAC9C,CAAC;IACD,MAAM,IAAI,KAAK,CACb,6FAA6F,CAC9F,CAAC;AACJ,CAAC;AAED,MAAM,KAAK,GAAG,IAAI,GAAG,EAAkB,CAAC;AAExC;;;;GAIG;AACH,MAAM,UAAU,YAAY,CAAC,IAAY;IACvC,MAAM,MAAM,GAAG,KAAK,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC;IAC/B,IAAI,MAAM,KAAK,SAAS;QAAE,OAAO,MAAM,CAAC;IACxC,MAAM,IAAI,GAAG,YAAY,CAAC,OAAO,CAAC,YAAY,EAAE,EAAE,IAAI,CAAC,EAAE,MAAM,CAAC,CAAC,OAAO,CAAC,KAAK,EAAE,EAAE,CAAC,CAAC;IACpF,KAAK,CAAC,GAAG,CAAC,IAAI,EAAE,IAAI,CAAC,CAAC;IACtB,OAAO,IAAI,CAAC;AACd,CAAC"}

package/dist/prompts/templates/dynamic-eval.md

@@ -0,0 +1,5 @@
+# Task — Run the harness dynamic evaluation
+
+Legacy entry point — agentrig now drives producer + judge separately via the
+scenario runner. Run `agentrig eval --dynamic` (which calls the new orchestrator)
+instead of relying on this prompt. Scope: {{SCOPE}}. Run id: {{RUN_ID}}.

package/dist/prompts/templates/investigate.md

@@ -0,0 +1,19 @@
+# Task 1 of 2 — Investigate this repository
+
+Investigate the repository in your current working directory and write your findings to
+`.agentrig/context.md` (create the file; create the .agentrig directory if needed).
+
+Cover, with evidence from actual files:
+1. **Purpose** — what this project is and who uses it (2-4 sentences).
+2. **Stack** — languages, frameworks, runtimes, package manager(s).
+3. **Commands** — the real install / build / test / lint commands (cite where you found them:
+   package.json scripts, Makefile, CI workflow, pyproject, go.mod, etc.). If a command does not
+   exist, say so explicitly rather than guessing.
+4. **Layout** — a concise directory map of the most important folders and what they contain.
+5. **Conventions** — notable coding conventions, testing patterns, and any "instructions are the
+   source of truth" docs (AGENTS.md, CONTRIBUTING, etc.).
+6. **Risks for an autonomous agent** — protected areas, generated code, flaky tests, anything an
+   agent should be careful about.
+
+Be thorough but factual. Do not modify any other files in this step. When done, reply with a short
+confirmation and the exact install/build/test/lint commands you found.

package/dist/prompts/templates/judge.md

@@ -0,0 +1,38 @@
+# Task — Score a completed scenario as an INDEPENDENT JUDGE
+
+You are the **judge** for scenario `{{SCENARIO}}` (type: `{{TYPE}}`). The producer
+agent has already finished. Read these files in your cwd to do your scoring:
+
+- `prompt.md`     — the exact task the producer was given
+- `diff.patch`    — the change the producer produced
+- `transcript.md` — the producer's own summary of what they did (BEWARE: don't be biased by it)
+- `oracle.json`   — deterministic axes (already scored — DO NOT re-score these)
+- `judge_brief.md` (if present) — calibration hints for soft axes only
+
+## What to score
+Score these soft axes against `{{RUBRIC_PATH}}`:
+{{AXES_LIST}}
+
+Tiers are strict: `0` / `0.5` / `1.0`. Any score < 1.0 MUST cite an issue code
+from that axis's registry plus a one-line evidence string. Use `confidence: 0` for
+axes you genuinely cannot observe.
+
+## How to submit
+Write your scores to `{{OUTPUT_JSON_PATH}}` in this exact shape:
+
+```json
+{
+  "axes": [
+    { "name": "self_verification", "score": 1.0, "confidence": 1 },
+    { "name": "clarity",           "score": 0.5, "confidence": 1, "code": "OQ-CLARITY-NAMING", "evidence": "function names use single letters" },
+    { "name": "memory",            "score": 0,   "confidence": 0 }
+  ]
+}
+```
+
+Do NOT save scores via `score.mjs` yourself — the orchestrator does that.
+
+## Independence
+Do NOT defer to the producer's reasoning. Decide each axis on the evidence in
+the diff + oracle results, not what the producer claims about their own work.
+If the diff contradicts the transcript, the diff wins.

package/dist/prompts/templates/producer-baseline-note.md

@@ -0,0 +1 @@
+**This is a BASELINE trial — harness OFF.** Do NOT read or follow `AGENTS.md`, `.agents/rules/`, `.agents/skills/`, or any AgentRig-installed instruction surface, even if they happen to be present in this worktree. Behave as a bare agent with only your training-data priors.

package/dist/prompts/templates/producer-handoff-checklist.md

@@ -0,0 +1,24 @@
+## Pre-handoff checklist (read before you reply)
+
+You are running with the AgentRig harness ON. Before declaring done, walk this checklist out loud
+in your transcript. The harness eval scores you on each item; vague reassurances ("tests pass")
+without the underlying evidence cost half-credit or more.
+
+- [ ] **Baseline captured.** Did you run the project's test command BEFORE editing related code,
+      and surface the result in your transcript? For a fix scenario: explicitly note the failing
+      test name and the error. For a feature scenario: note the suite was green.
+      *Bad:* "All tests pass."
+      *Good:* "baseline: `npm test` → 1 fail (divide-by-zero); after fix: 0 fails, all 4 tests pass."
+
+- [ ] **After captured.** Did you re-run the full test command at the end and surface the new
+      state? The transition baseline → after is the evidence that your edit did what you claim.
+
+- [ ] **Wiki entry committed for any non-obvious lesson.** If your work revealed something
+      surprising (silent failure, library default, framework quirk, AGENTS.md rule that almost
+      bit you), use the `log-gotcha` skill to write a `.agents/wiki/<topic>.md` entry IN THE
+      SAME DIFF. Acknowledging the lesson only in your summary is half-credit. Silent is zero.
+      Run `git diff --cached --stat` to confirm the wiki file is staged.
+
+- [ ] **Diff is on-target.** `git diff --stat` should show only files you intentionally changed.
+
+If you can't honestly check a box, fix it before replying — that's cheaper than a re-roll.

package/dist/prompts/templates/producer-harness-note.md

@@ -0,0 +1 @@
+**This is a HARNESS trial — harness ON.** Follow `AGENTS.md`, the rules in `.agents/rules/`, and the skills in `.agents/skills/` if they are present in this worktree.

package/dist/prompts/templates/producer.md

@@ -0,0 +1,10 @@
+# Scenario task
+{{VARIANT_NOTE}}
+Your entire job is described below. Work inside the current directory (this is a
+throwaway worktree dedicated to your trial). When done, simply finish — the
+scenario runner captures your diff, your transcript, and runs the deterministic
+oracle automatically.
+
+---
+
+{{SCENARIO_PROMPT}}{{HANDOFF_CHECKLIST}}

package/dist/prompts/templates/scaffold-scenarios.md

@@ -0,0 +1,56 @@
+# Task — Generate repository-specific eval scenarios
+
+The 3 scenarios under `.agentrig/eval/scenarios/` are language-agnostic JS micro-fixtures. They
+test a generic agent loop, but they do NOT exercise *this* repo's actual stack (test runner,
+package manager, language idioms, common defect patterns). Your job: write {{COUNT}} new
+scenario(s) that ARE specific to this repo.
+
+## Repo investigation (from `.agentrig/context.md`)
+
+```
+{{CONTEXT_MD}}
+```
+
+## What a scenario looks like (templates)
+
+{{EXAMPLES_TEXT}}
+
+## What to produce
+
+For each new scenario:
+
+1. Create a directory `.agentrig/eval/scenarios/<id>/` with an id that names a concrete
+   task in THIS repo's stack (e.g. `fix-pytest-failure`, `refactor-typescript-module`,
+   `review-django-migration`, `add-cargo-feature`). NO generic ids — `fix-failing-test` is taken.
+2. Write `scenario.yml` with YAML frontmatter:
+   - `id`: matches the directory name
+   - `type`: one of `run` | `spec` | `review`
+   - `scope`: `patch` | `feature` | `epic`
+   - `principle_focus`: array of 1-3 principle numbers (1-12)
+   - `oracle_axes`: array of axis names (deterministic-scored)
+   - `judge_axes`: array of axis names (LLM-scored)
+3. Write `prompt.md` — the exact task handed to the producer agent. NO ambiguity, NO "invent your own spec."
+4. Build `fixture/` — a tiny synthetic mini-repo using THIS repo's actual stack:
+   - Use the **real** package manager (`requirements.txt` / `go.mod` / `package.json` / `Cargo.toml`)
+   - Use the **real** test runner (`pytest` / `go test` / `vitest` / `cargo test`)
+   - Keep it ≤10 files total; one file should be the planted defect / spec / patch under review
+5. Write `oracle.yml` — deterministic checks (cmd, diff_stats, diff_files, file_contains, file_missing).
+   The `cmd` checks MUST use this repo's actual test command, not `npm test`.
+6. Write `README.md` — 1-2 paragraphs describing what the scenario tests + what a defect looks like.
+7. Write `judge_brief.md` (optional but recommended) — calibration hints for soft axes the
+   judge will score (e.g. "1.0 = wrote a wiki entry, 0.5 = mentioned in summary, 0 = silent").
+
+## Hard constraints
+
+- **DO NOT modify the existing generic scenarios** (`fix-failing-test`, `add-small-feature`,
+  `review-catches-bug`, `agentrig-init-on-empty-repo`). They stay as both templates AND running scenarios.
+- **DO NOT touch any file outside `.agentrig/eval/scenarios/`.**
+- **Axis names must come from the live registry.** Valid types: {{AXIS_TYPES}}.
+  Valid axis names (use only these): {{AXIS_NAMES}}.
+- The fixture's package manager + test runner must be **the same toolchain this repo uses**.
+  Check `AGENTS.md` for the install/test commands.
+- Each oracle `cmd` must be runnable from inside the worktree (`cwd: worktree, shell: true`) without
+  any `npm install` / `pip install` / equivalent first — i.e., the fixture should be self-contained
+  or rely on stdlib only. If the test command needs deps, include a tiny dependency-free alternative.
+
+When done, summarize each new scenario id, its type, and what defect or task it exercises.

package/dist/prompts/templates/system-message.md

@@ -0,0 +1,5 @@
+You are AgentRig's installer agent. You are setting up and tailoring an autonomous-coding-agent
+"harness" inside a real repository. Be precise and surgical. Prefer reading many files to understand
+the repo before writing. Only edit the files you are explicitly asked to. Never invent build/test
+commands — verify them against package manifests, lockfiles, CI config, and scripts you actually
+find. Keep edits idempotent.

package/dist/prompts/templates/tailor.md

@@ -0,0 +1,38 @@
+# Task 2 of 2 — Tailor the installed harness to this repository
+
+AgentRig has just installed a canonical best-practice harness into this repo:
+
+{{ARTIFACT_LIST}}
+
+Using everything you learned in Task 1 (and `.agentrig/context.md`), make the harness
+repo-specific. Edit **only** the files listed below.
+
+1. **`AGENTS.md`** — replace every `{{PLACEHOLDER}}` and fill the content between the
+   `<!-- AGENTRIG:...:start -->` / `:end` markers:
+   - `{{REPO_NAME}}`, `{{REPO_SUMMARY}}` — name and a 2-3 sentence description.
+   - The `commands` block — the REAL install/build/test/lint commands you verified. If one
+     genuinely does not exist, write `(none)`.
+   - The `dirmap` block — a concise directory map.
+   Do NOT change anything between the `critical-rules` markers.
+2. **`.agents/rules/coding-standards.md`** — replace the generic baseline with standards that
+   actually match this repo's language and conventions. Keep it to a short list of imperative
+   reflexes and keep the frontmatter `globs`/`description`.
+3. **`.agentrig/eval/scenarios/`** — adjust the existing scenario files so the setup/success
+   criteria reference this repo's real test/build commands and structure. Do not remove the axis
+   lists.
+4. **`.github/workflows/copilot-setup-steps.yml`** — author a REAL, repo-specific setup workflow so
+   the GitHub Copilot **cloud/coding agent** has a ready environment (don't leave a generic stub).
+   Base it on your investigation:
+   - A single job named EXACTLY `copilot-setup-steps` on `runs-on: ubuntu-latest`, with
+     `permissions: contents: read`, triggered by `workflow_dispatch` + `push`/`pull_request`
+     filtered to this file.
+   - Steps that install the ACTUAL toolchain + dependencies you found: correct language runtime(s)
+     and version(s) (from `.nvmrc`/`.tool-versions`/`engines`/`go.mod`/`pyproject.toml`), the
+     correct package manager and install command (e.g. `npm ci`/`pnpm i --frozen-lockfile`/
+     `pip install -e .`/`go mod download`), dependency caching, and any system packages or
+     `services` (databases, etc.) the build/tests need. Keep it to env setup — not the task itself.
+   If you cannot determine the stack confidently, leave the generated scaffold and note what's
+   missing.
+
+Keep all YAML frontmatter and the AgentRig markers intact. Do not touch the state machine, role
+files, MCP config, or the eval scripts. When finished, summarize exactly which files you changed.

package/dist/prompts/templates/update.md

@@ -0,0 +1,14 @@
+# Task — Re-apply the latest AgentRig best practices
+
+AgentRig refreshed these canonical artifacts to their latest version:
+{{CHANGED_LIST}}
+
+For each refreshed file, reconcile it with this repo:
+- Preserve repo-specific content the team added (especially inside AgentRig markers in AGENTS.md and
+  in `coding-standards.md` and the scenarios).
+- Adopt new structure, new sections, and new defaults from the canonical version.
+- If there is a genuine conflict, prefer the new canonical structure but keep repo-specific facts
+  (commands, directory map, summary).
+
+Re-read `.agentrig/context.md` first for repo context. Summarize what you merged and any conflicts
+you resolved.

package/knowledge/manifest.json

@@ -1,6 +1,6 @@
 {
   "$schema": "agentrig-manifest/1",
-  "knowledgeVersion": "0.6.0",
+  "knowledgeVersion": "0.6.1",
   "description": "Declares which best-practice artifacts AgentRig installs into a target repo and where. `src` is relative to the knowledge/ root; `dest` is relative to the target repo root. `kind`: file | dir | template. Templates contain {{PLACEHOLDERS}} the agent fills from its investigation; deterministic installs substitute known values and leave the rest for the agent.",
   "artifacts": [
     {

package/knowledge/templates/agents/README.md

@@ -8,10 +8,10 @@ single-model-bias mitigation surfaces problems no single model would catch alone
 
 | Role | File | Default model | Drives state |
 |------|------|---------------|--------------|
-| **triager**  | `triager.{yml,md}`  | `gpt-5.5` (high)          | `ingested → queued` |
-| **developer**| `developer.{yml,md}`| `claude-opus-4.8` (high)  | `queued → implementing → reviewing` |
-| **reviewer** | `reviewer.{yml,md}` | `gpt-5.5` (high)          | `reviewing` |
-| **judge**    | `judge.{yml,md}`    | `claude-opus-4.8` (high)  | `judging → ready_to_merge` |
+| **triager**  | `triager.{yml,md}`  | `gpt-5.5` (premium)          | `ingested → queued` |
+| **developer**| `developer.{yml,md}`| `claude-opus-4.8` (premium)  | `queued → implementing → reviewing` |
+| **reviewer** | `reviewer.{yml,md}` | `gpt-5.5` (premium)          | `reviewing` |
+| **judge**    | `judge.{yml,md}`    | `claude-opus-4.8` (premium)  | `judging → ready_to_merge` |
 
 > Keep the **reviewer on a different model family than the developer**. The audit
 > (`agentrig eval --static`) checks for this.

package/knowledge/templates/agents/developer.yml

@@ -1,7 +1,7 @@
 # Developer role (principle 2). Implements the change in the `implementing` state.
 role: developer
 model: claude-opus-4.8
-model_tier: high
+model_tier: premium
 # Skills are auto-discovered from .agents/skills; no explicit list needed.
 allowed_tools: [read, write, edit, bash, grep, glob]
 prompt: agents/developer.md

package/knowledge/templates/agents/judge.yml

@@ -1,6 +1,6 @@
 # Judge role (principle 2, 6). Independent, rubric-driven scoring before merge.
 role: judge
 model: claude-opus-4.8
-model_tier: high
+model_tier: premium
 allowed_tools: [read, grep, glob, bash]
 prompt: agents/judge.md

package/knowledge/templates/agents/reviewer.yml

@@ -2,6 +2,6 @@
 # to mitigate single-model bias — divergent verdicts surface problems neither model alone catches.
 role: reviewer
 model: gpt-5.5
-model_tier: high
+model_tier: premium
 allowed_tools: [read, grep, glob, bash]
 prompt: agents/reviewer.md

package/knowledge/templates/agents/triager.yml

@@ -4,6 +4,6 @@
 # routine routing and would be a sensible cost optimization if triage volume scales up.
 role: triager
 model: gpt-5.5
-model_tier: high
+model_tier: premium
 allowed_tools: [read, grep, glob, bash]
 prompt: agents/triager.md

package/knowledge/templates/rules/coding-standards.md

@@ -1,6 +1,7 @@
 ---
 globs: ["**/*"]
 description: Baseline coding standards applied to every change in this repo.
+priority: 3
 ---
 
 # Coding standards (reflex)

package/knowledge/templates/skills/harness-eval/SKILL.md

@@ -1,6 +1,10 @@
 ---
 name: harness-eval
 description: Evaluate THIS repository's agent harness — a deterministic structure audit (A1) plus content quality probes (A2), plus an isolated producer/judge dynamic eval (B) with paired sign-test A/B variant comparison.
+triggers:
+  - "evaluate / score the harness (static or dynamic)"
+  - "did a harness change improve or regress it?"
+  - before merging changes to skills/rules/agents/prompts
 allowed-tools: Bash Read Grep Glob
 argument-hint: "[--static|--dynamic] [--scenario id] [--variant v] [--n trials]"
 ---

package/knowledge/templates/skills/skill-authoring/SKILL.md

@@ -4,7 +4,7 @@ description: Admission bar and structure for writing a new skill, so the skill l
 triggers:
   - "create / add a new skill"
   - "this procedure keeps coming up"
-allowed-tools: Read Edit Grep Glob
+allowed-tools: Read Write Edit Grep Glob
 argument-hint: "<skill-name>"
 ---
 

package/knowledge/templates/skills/skill-improver/SKILL.md

@@ -4,7 +4,7 @@ description: Turn a reviewer/judge failure into an instruction-surface change th
 triggers:
   - "a mistake recurred"
   - "reviewer feedback points at a missing rule/skill"
-allowed-tools: Read Edit Grep Glob
+allowed-tools: Read Write Edit Grep Glob
 argument-hint: "<short description of the failure>"
 ---
 

package/knowledge/templates/wiki/README.md

@@ -13,11 +13,7 @@ Before adding an entry, confirm no existing entry covers it. If one does, **shar
 adding a near-duplicate. Each entry should be: a title, the symptom, the root cause, the fix, and a
 one-line prevention.
 
-## Entry template
-```markdown
-### <short title>
-- **Symptom:** what went wrong / how it showed up
-- **Cause:** the real root cause
-- **Fix:** the change that resolved it
-- **Prevention:** the rule/skill wording that would have stopped it (feed to skill-improver)
-```
+## Adding an entry
+Copy `_TEMPLATE.md` to `.agents/wiki/<slug>.md`, fill its four sections, and add a one-line link to
+the **Index** in `index.md` (newest first). See `index.md` for what belongs in the wiki vs a skill,
+rule, or `AGENTS.md`.

package/knowledge/templates/wiki/index.md

@@ -15,15 +15,10 @@ way. It is **not** a mirror of the docs or skills.
 If a gotcha becomes a reusable procedure, **promote it to a skill** and leave a one-line pointer
 here.
 
-## Tiers (principle 8)
-1. **Central wiki (this dir, committed):** repo-wide, reviewed gotchas. CODEOWNERS-gate it.
-2. **Local wiki (git-ignored `*.local.md`):** machine/contributor-specific notes.
-3. **Session scratch:** ephemeral working notes; never a substitute for the wiki.
-
 ## Index
 _Add a one-line link per entry as you create it, newest first._
 - (none yet)
 
-## Admission test (strict — duplication kills wikis)
-Before adding an entry, confirm no existing entry covers it. If one does, **sharpen it** instead of
-adding a near-duplicate. Use the format in `_TEMPLATE.md`.
+---
+Wiki **policy** (tiers + admission test) lives in `README.md`; the **entry format** lives in
+`_TEMPLATE.md`. Don't restate them here.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@doidor/agentrig",
-  "version": "0.11.0",
+  "version": "0.11.1",
   "description": "AgentRig — an agentic meta-harness. A CLI that investigates a repository and installs (and evaluates) a best-practice agent harness.",
   "type": "module",
   "bin": {
@@ -27,7 +27,7 @@
     "node": ">=22.0.0"
   },
   "scripts": {
-    "build": "tsc -p tsconfig.json",
+    "build": "tsc -p tsconfig.json && node scripts/copy-prompt-templates.mjs",
     "dev": "tsc -p tsconfig.json --watch",
     "clean": "rm -rf dist",
     "prepare": "npm run build",