darwin-agents 0.5.0-alpha.1 → 0.5.0-alpha.2

package/CHANGELOG.md

@@ -1,136 +1,89 @@
 # Changelog
 
-## [0.5.0-alpha.1] — 2026-05-24
+## [0.5.0-alpha.2] — 2026-05-25
 
-**Phase 2 A1: Execution-Trace-Capture.** First pre-release of Darwin's
-Phase 2 tech roadmap. Unblocks GEPA-style reflective optimizers (A2)
-and validate-by-reproduce drift-detection (A5) by giving them a
-structured trajectory to consume.
-
-Industry-aligned with the 2026 agent-observability consensus (Braintrust,
-Langfuse, Strands SDK, Microsoft Foundry, OTEL GenAI semantic conventions):
-three span types — Tool / Reasoning / Turn-level errors — captured into a
-single `ExecutionTrace` object, persisted as JSONB (Postgres) or TEXT
-(SQLite), and tagged with a forward-compatible `version: 1` discriminator.
+**GEPA-Style Reflective Optimizer (Phase 2 A2).** Multi-objective Pareto
+selection + text-feedback-driven prompt mutation as a TS-native
+adaptation of the GEPA framework (arxiv 2507.19457). Released under the
+`alpha` npm dist-tag in parallel with v0.5.0-alpha.1 (execution-trace
+capture, A1). `npm install darwin-agents@alpha` resolves to
+0.5.0-alpha.2; `npm install darwin-agents` stays on 0.4.9 (latest).
 
 ### Added
 
-- **`ExecutionTrace` schema** (`src/types.ts`) — versioned trajectory shape:
-  `toolCalls[]` (with OTEL-mappable `id` / `tool` / `args` / `resultSummary`
-  (2000-char cap) / `outcome` / `durationMs` / `retryCount?` / `errorClass?` /
-  `errorMessage?` / `turn`), `textBlockCount` (honest name — NOT a thinking-
-  block counter, V2 will add typed `reasoningBlocks`), `turnCount`,
-  `mcpInvocations`, `errors[]` (turn-level), `tokenUsage?` (OTEL `gen_ai.usage.*`
-  fields: input/output/cache_read/cache_creation tokens), `capturedAt`. Plus
-  optional `trajectory?: ExecutionTrace` on `DarwinExperiment` (additive —
-  pre-A1 callers unaffected).
-
-- **`createTraceCapture()` factory** (`src/core/trace-capture.ts`) — pure,
-  transport-agnostic capturer. The runtime feeds tool events; the capturer
-  aggregates into a typed trajectory. API:
-
-  ```ts
-  const trace = createTraceCapture();
-  trace.startTurn();
-  trace.recordToolUse('toolu_01AB', 'mcp__nex__search', { query: 'x' });
-  trace.recordToolResult('toolu_01AB', 'success', { resultSummary: '3 hits' });
-  trace.recordTextBlock();
-  trace.addTokens({ inputTokens: 1200, outputTokens: 340 });
-  trace.recordError('parse_error', 'invalid JSON');
-  const trajectory = trace.finalize();
-  ```
-
-  Unpaired `recordToolUse` calls (no matching `recordToolResult` before
-  `finalize`) surface as `outcome: 'error', errorClass: 'unpaired_call'`
-  so silent SDK hangs remain visible in the trace. Customizable via
-  `TraceCaptureOptions`: `now?` (clock injection for tests),
-  `isMcpTool?` (predicate override for non-`mcp__`-prefixed servers).
-
-- **`addTokens()` aggregator** — lossy-merge of per-turn LLM usage. Missing
-  fields (`NaN` / `Infinity` / `undefined`) skip silently rather than
-  defaulting to zero — preserves the distinction between "provider didn't
-  report" and "actually zero tokens".
-
-- **JSONB persistence** in `darwin_experiments.trajectory` column +
-  `idx_darwin_exp_trajectory_gin` GIN index (Postgres) for `@>`
-  containment queries from A2 / A5 consumers. SQLite stores the same
-  shape as JSON-stringified TEXT.
-
-- **`scripts/migrate-add-trajectory.ts`** — idempotent migration script.
-  Pre-checks column + index existence (filtered by `current_schema()`
-  for multi-schema-safe operation), runs `ALTER TABLE … ADD COLUMN IF
-  NOT EXISTS trajectory JSONB` + `CREATE INDEX IF NOT EXISTS`, then
-  verifies. Rollback path documented inline.
-
-  ```bash
-  DARWIN_POSTGRES_URL=postgresql://… npx tsx scripts/migrate-add-trajectory.ts
-  ```
-
-- **Defensive parsing** in both memory backends — `parseTrajectory` /
-  `parseTrajectoryColumn` drop malformed values (wrong `version`,
-  non-object, invalid JSON) to `undefined` instead of crashing the
-  load. Future schema versions (`version !== 1`) are silently ignored
-  so v0.5 consumers don't break on v0.6 trajectories.
-
-- **39 new tests** across two suites (all green):
-  - `tests/trace-capture.test.ts` (32 unit tests): basic flow,
-    defensive behaviour, truncation (2000-char `resultSummary`),
-    MCP-heuristic, schema invariants, tool_call_id passthrough,
-    `addTokens` aggregate semantics
-  - `tests/memory-trajectory.test.ts` (7 tests): SQLite roundtrip,
-    backward-compat with pre-A1 rows, defensive parsing, idempotent
-    migration, Postgres-gated JSONB roundtrip
-
-### Changed
-
-- **DDL single-source-of-truth** — the trajectory column is defined
-  ONLY in the additive `ALTER TABLE … ADD COLUMN IF NOT EXISTS` path
-  (Postgres) / PRAGMA-guarded ALTER (SQLite), never inline in the
-  `CREATE TABLE`. Schema-evolution lives in one place; fresh installs
-  reach the same end-state as legacy installs.
-
-- **Postgres `ON CONFLICT` preserves trajectory** on feedback-only
-  re-saves via `COALESCE(EXCLUDED.trajectory, darwin_experiments.trajectory)`.
-  This means a second `saveExperiment(exp)` call that omits trajectory
-  doesn't zero out the previously-stored trace.
-
-  **NOTE — SQLite asymmetry:** SQLite uses `INSERT OR REPLACE` which
-  drops + re-inserts the row, so callers wanting to preserve a prior
-  trajectory across re-saves MUST include it in the new payload. This
-  asymmetry is documented on `MemoryProvider.saveExperiment` in the
-  interface JSDoc.
-
-### Backwards compatibility
-
-100% backwards-compatible. The new `trajectory` field is optional, the
-new column is nullable, the new methods on `MemoryProvider` are
-additive. Existing v0.4.x consumers see no behavioural changes.
-
-Verified on a live `darwin_db` with 341 experiments, 339 of which
-pre-date A1 — all loaded cleanly with `trajectory: undefined`.
-
-### Why "alpha.1"
-
-`textBlockCount` is honest but limited — V2 will replace it with a
-typed `reasoningBlocks: ReasoningBlock[]` sequence carrying the actual
-text content per reasoning step, which is what GEPA reflectors need
-for per-decision blame attribution. Existing `textBlockCount` will stay
-as a fast aggregate. The `alpha.1` tag signals the schema is subject to
-this kind of additive evolution before `0.5.0` final.
-
-Three known minor gaps (deferred to follow-up patches):
-
-- Per-call cost attribution (token usage per tool invocation, not just
-  per-run aggregate)
-- Trace-capture lazy-load flag stays permanent on transient import
-  failure (low impact: Darwin is either built or not)
-- Token extraction in the SDK adapter is Anthropic-shaped (`message.usage`)
-  and may silently miss tokens for non-Anthropic providers — by design
-  (token usage is documented optional), but a debug-level log line in a
-  follow-up patch will make this easier to spot.
-
-Install: `npm install darwin-agents@alpha`. The default `latest` tag
-remains on `0.4.9` until `0.5.0` final ships.
+- **`GepaOptimizer`** — generation-loop wrapper producing N variant
+  mutations per call (default N=3, [1, 10]). Three `feedbackStrategy`
+  modes: `"split"` (round-robin partition, diversity), `"replicate"`
+  (every variant sees all feedback), `"single"` (one reflection).
+  Separate `nextGeneration(scored, opts)` Pareto-selects survivors for
+  the next generation.
+- **`Reflector`** — single-shot LLM call with GEPA's "smallest possible
+  targeted edit" template. Output is cleaned (fences stripped) and
+  truncated at sentence boundary.
+- **`pareto.ts`** — `dominates` / `nonDominatedFront` / `paretoSelect` /
+  `scalarise` pure helpers + `DARWIN_DEFAULT_OBJECTIVES` constant
+  (matching `DarwinMetrics` field names + existing weight scheme).
+- **`RunPromptFn`** — shared injected-LLM-call type, single source of
+  truth for both `PromptOptimizer` and `Reflector`.
+- **A1 sync (S1184):** `createTraceCapture` + `ExecutionTrace` /
+  `TraceToolCall` / `TraceTokenUsage` / `TraceTurnError` now exported
+  from the OS package (were already in v0.5.0-alpha.1 on npm, OS source
+  catches up this release).
+
+### Deliberate deviations from GEPA paper (documented in source)
+
+- N variants per `generate()` call vs GEPA Algorithm 1's 1-offspring-
+  per-iteration.
+- `feedbackStrategy: "split"` is our adaptation, not in the paper.
+- `paretoSelect` truncation uses scalarised tie-break, not GEPA
+  Algorithm 2's coverage-proportional sampling — V0.6 will add
+  `truncationStrategy: "coverage" | "crowding"`.
+- GEPA+Merge (paper Appendix F, ~+5% lift) NOT implemented — V0.6.
+- Instance-wise coverage sampling NOT implemented — V0.6.
+- Single injected `runPrompt` for both task and reflection — GEPA docs
+  recommend stronger `reflection_lm`. Optional `reflectionRunPrompt`
+  override deferred to V0.5.1.
+
+### Fixed (R1 + R2 V0.5.0-alpha.2 code-review findings)
+
+The 3-Agent code-review loop ran twice. R1 found 13 findings, R2 caught
+2 must-fix that R1 missed. All addressed pre-publish.
+
+**R1 — 6 MUST-FIX (S1185):**
+
+1. **HIGH (Critic H1):** Template injection — `String.replace` order
+   meant `currentPrompt` containing `{FEEDBACKS}` literal could trigger
+   double-substitution. Fixed by substituting `{CURRENT_PROMPT}` last.
+2. **HIGH (Critic H2):** `feedbackCap` accepted negative values — added
+   `Math.max(1, Math.floor(...))` guard.
+3. **HIGH (Analyst A5):** `ParetoObjective` JSDoc example used wrong
+   `DarwinMetrics` field names. Fixed + `DARWIN_DEFAULT_OBJECTIVES`
+   constant.
+4. **HIGH (Analyst A1):** `RunPromptFn` was duplicated. Extracted to
+   `evolution/run-prompt-fn.ts`.
+5. **MED (Critic M2):** `nextGeneration` used reference-identity on
+   `metrics` — switched to explicit index-based mapping (refactor-safe).
+6. **MED (Critic M4):** Added scale-normalization JSDoc warning on
+   `ParetoObjective.weight`.
+
+**R2 — 2 MUST-FIX (caught what R1 missed, S1185):**
+
+7. **CRITICAL (R2-C1):** R1's clamp `Math.max(1, Math.floor(NaN)) ===
+   NaN` — silent bypass for NaN/Infinity. Hardened with
+   `Number.isFinite()` + fallback to default.
+8. **LOW (R2-L1):** `generate("p", [])` threw opaque internal error.
+   Added GEPA-specific boundary validation pointing callers at
+   `PromptOptimizer` for cold-start. Plus R2-M1 guard for shared
+   metrics-object references.
+
+### Test coverage
+
+- **307/308 OS tests green** (1 pre-existing skip, 0 fail). Was 268 in
+  v0.4.9. New test files: `pareto.test.ts` (16), `reflector.test.ts`
+  (14), `optimizer-gepa.test.ts` (12), `r1-fixes.test.ts` (12 R1+R2
+  regression). A1 trace + memory-trajectory tests synced from
+  v0.5.0-alpha.1.
+- tsc strict + build clean.
 
 ## [0.4.9] — 2026-05-22
 

package/dist/agents/analyst.d.ts

@@ -0,0 +1,11 @@
+/**
+ * Analyst Agent — Code Intelligence
+ *
+ * Analyzes codebases for quality, patterns, security issues,
+ * and improvement opportunities.
+ *
+ * Uses filesystem access (Read, Glob, Grep tools).
+ */
+import type { AgentDefinition } from '../types.js';
+export declare const analyst: AgentDefinition;
+//# sourceMappingURL=analyst.d.ts.map

package/dist/agents/analyst.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"analyst.d.ts","sourceRoot":"","sources":["../../src/agents/analyst.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAEH,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,aAAa,CAAC;AAEnD,eAAO,MAAM,OAAO,EAAE,eAqErB,CAAC"}

package/dist/agents/analyst.js

@@ -0,0 +1,78 @@
+/**
+ * Analyst Agent — Code Intelligence
+ *
+ * Analyzes codebases for quality, patterns, security issues,
+ * and improvement opportunities.
+ *
+ * Uses filesystem access (Read, Glob, Grep tools).
+ */
+export const analyst = {
+    name: 'analyst',
+    role: 'Code Analyst',
+    description: 'Analyzes codebases for quality, patterns, and issues. Finds what humans miss.',
+    maxTurns: 25,
+    tools: ['Read', 'Glob', 'Grep', 'Bash'],
+    systemPrompt: `You are a senior code analyst who reviews codebases for quality and issues.
+
+YOUR MISSION:
+Analyze the given codebase or file path and deliver a structured quality report.
+
+ANALYSIS PROCESS:
+1. Scan the project structure (Glob for key files: package.json, tsconfig, etc.)
+2. Read key files to understand architecture
+3. Search for common issues (Grep for patterns)
+4. Evaluate code quality, security, and architecture
+
+WHAT TO LOOK FOR:
+- **Architecture**: Project structure, dependency management, module boundaries
+- **Code Quality**: TypeScript strictness, error handling, naming conventions
+- **Security**: Hardcoded secrets, injection risks, unsafe patterns
+- **Performance**: N+1 queries, missing indexes, unnecessary re-renders
+- **Dead Code**: Unused exports, unreachable branches, commented-out code
+- **Dependencies**: Outdated packages, known vulnerabilities, bundle size
+
+SEVERITY LEVELS:
+- P0 (Critical): Security vulnerabilities, data loss risks
+- P1 (High): Bugs, performance issues, architectural problems
+- P2 (Medium): Code quality, maintainability issues
+- P3 (Low): Style, naming, minor improvements
+
+RULES:
+- Only report issues you can PROVE (show the file and line)
+- Don't nitpick formatting — focus on substance
+- Prioritize by impact, not by count
+- Suggest concrete fixes, not vague advice
+- If the code is good, say so — don't invent problems
+
+OUTPUT FORMAT:
+# Code Analysis: {Project/Path}
+
+## Summary
+{2-3 sentences: overall health, biggest concern, biggest strength}
+
+## Critical Issues (P0-P1)
+### {Issue Title}
+- **File**: {path}:{line}
+- **Problem**: {specific description}
+- **Fix**: {concrete suggestion}
+
+## Improvements (P2-P3)
+- {issue + file + suggestion}
+
+## Architecture Notes
+{Observations about structure, patterns, decisions}
+
+## Score: {1-10}/10`,
+    evolution: {
+        enabled: true,
+        evaluator: 'critic',
+        metrics: {
+            quality: 0.45,
+            sourceCount: 0.05,
+            outputLength: 0.15,
+            duration: 0.10,
+            success: 0.25,
+        },
+    },
+};
+//# sourceMappingURL=analyst.js.map

package/dist/agents/analyst.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"analyst.js","sourceRoot":"","sources":["../../src/agents/analyst.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAIH,MAAM,CAAC,MAAM,OAAO,GAAoB;IACtC,IAAI,EAAE,SAAS;IACf,IAAI,EAAE,cAAc;IACpB,WAAW,EAAE,+EAA+E;IAC5F,QAAQ,EAAE,EAAE;IACZ,KAAK,EAAE,CAAC,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,CAAC;IACvC,YAAY,EAAE;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;oBAkDI;IAElB,SAAS,EAAE;QACT,OAAO,EAAE,IAAI;QACb,SAAS,EAAE,QAAQ;QACnB,OAAO,EAAE;YACP,OAAO,EAAE,IAAI;YACb,WAAW,EAAE,IAAI;YACjB,YAAY,EAAE,IAAI;YAClB,QAAQ,EAAE,IAAI;YACd,OAAO,EAAE,IAAI;SACd;KACF;CACF,CAAC"}

package/dist/agents/blog-writer.d.ts

@@ -0,0 +1,13 @@
+/**
+ * Blog Writer Agent — SEO-Optimized Content
+ *
+ * Writes SEO-optimized blog posts with keyword focus.
+ * No MCP tools — pure text generation.
+ * Blog-specific critics: SEO, readability, conversion.
+ *
+ * Customize brand/site by providing context in the task prompt
+ * or by creating a custom agent with defineAgent().
+ */
+import type { AgentDefinition } from '../types.js';
+export declare const blogWriter: AgentDefinition;
+//# sourceMappingURL=blog-writer.d.ts.map

package/dist/agents/blog-writer.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"blog-writer.d.ts","sourceRoot":"","sources":["../../src/agents/blog-writer.ts"],"names":[],"mappings":"AAAA;;;;;;;;;GASG;AAEH,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,aAAa,CAAC;AAEnD,eAAO,MAAM,UAAU,EAAE,eAgDxB,CAAC"}

package/dist/agents/blog-writer.js

@@ -0,0 +1,59 @@
+/**
+ * Blog Writer Agent — SEO-Optimized Content
+ *
+ * Writes SEO-optimized blog posts with keyword focus.
+ * No MCP tools — pure text generation.
+ * Blog-specific critics: SEO, readability, conversion.
+ *
+ * Customize brand/site by providing context in the task prompt
+ * or by creating a custom agent with defineAgent().
+ */
+export const blogWriter = {
+    name: 'blog-writer',
+    role: 'SEO Blog Writer',
+    description: 'Writes SEO-optimized blog posts. Keyword-aware, structured for readability and conversion.',
+    maxTurns: 8,
+    systemPrompt: `You are a senior SEO content writer for a premium digital agency.
+
+BLOG STANDARDS:
+- Write in the language the user specifies (default: English)
+- Target audience: SMB owners considering digital solutions, website redesign, or new projects
+- Tone: Expert but accessible. Explain tech concepts simply. No jargon without explanation.
+- Length: 800-1500 words unless specified otherwise
+
+SEO RULES:
+- Include the main keyword in: Title, first paragraph, one H2, meta description
+- Use 3-5 H2 headers with keyword variations (not exact stuffing)
+- Write a compelling meta description (max 155 chars)
+- Suggest internal link opportunities where relevant
+- Use short paragraphs (max 3-4 sentences)
+- Include a FAQ section with 3 questions (structured data opportunity)
+
+CONTENT QUALITY:
+- Lead with the reader's problem, not the solution
+- Include at least one concrete example, case study, or data point
+- Address objections ("But what about...") proactively
+- Every section must answer "why should I care?"
+- End with clear CTA (contact, consultation, related post)
+- NEVER invent statistics. Use "typically", "in our experience" for estimates.
+
+OUTPUT FORMAT:
+1. Title (H1) — compelling, keyword-included, under 60 chars
+2. Meta Description — under 155 chars
+3. Main content with H2 headers
+4. FAQ section (3 questions)
+5. CTA paragraph
+6. Suggested internal links`,
+    evolution: {
+        enabled: true,
+        evaluator: 'multi-critic',
+        metrics: {
+            quality: 0.55,
+            sourceCount: 0.0,
+            outputLength: 0.15,
+            duration: 0.05,
+            success: 0.25,
+        },
+    },
+};
+//# sourceMappingURL=blog-writer.js.map

package/dist/agents/blog-writer.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"blog-writer.js","sourceRoot":"","sources":["../../src/agents/blog-writer.ts"],"names":[],"mappings":"AAAA;;;;;;;;;GASG;AAIH,MAAM,CAAC,MAAM,UAAU,GAAoB;IACzC,IAAI,EAAE,aAAa;IACnB,IAAI,EAAE,iBAAiB;IACvB,WAAW,EAAE,4FAA4F;IACzG,QAAQ,EAAE,CAAC;IACX,YAAY,EAAE;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;4BA8BY;IAE1B,SAAS,EAAE;QACT,OAAO,EAAE,IAAI;QACb,SAAS,EAAE,cAAc;QACzB,OAAO,EAAE;YACP,OAAO,EAAE,IAAI;YACb,WAAW,EAAE,GAAG;YAChB,YAAY,EAAE,IAAI;YAClB,QAAQ,EAAE,IAAI;YACd,OAAO,EAAE,IAAI;SACd;KACF;CACF,CAAC"}

package/dist/agents/critic.d.ts

@@ -0,0 +1,11 @@
+/**
+ * Critic Agent — The Heart of Darwin
+ *
+ * Evaluates other agents' output. Without the Critic,
+ * there's no quality score, no evolution, no improvement.
+ *
+ * Zero-config: no MCP servers, no API keys.
+ */
+import type { AgentDefinition } from '../types.js';
+export declare const critic: AgentDefinition;
+//# sourceMappingURL=critic.d.ts.map

package/dist/agents/critic.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"critic.d.ts","sourceRoot":"","sources":["../../src/agents/critic.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAEH,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,aAAa,CAAC;AAEnD,eAAO,MAAM,MAAM,EAAE,eAgDpB,CAAC"}

package/dist/agents/critic.js

@@ -0,0 +1,57 @@
+/**
+ * Critic Agent — The Heart of Darwin
+ *
+ * Evaluates other agents' output. Without the Critic,
+ * there's no quality score, no evolution, no improvement.
+ *
+ * Zero-config: no MCP servers, no API keys.
+ */
+export const critic = {
+    name: 'critic',
+    role: "Devil's Advocate & Quality Evaluator",
+    description: "Reviews and scores other agents' output. Enables Darwin evolution through quality feedback.",
+    maxTurns: 5,
+    systemPrompt: `You are a sharp, constructive critic who evaluates AI agent outputs.
+
+YOUR ROLE:
+Score the quality of an agent's output on a scale of 1-10 and provide specific feedback.
+
+EVALUATION CRITERIA:
+1. **Accuracy** (0-10): Are claims factual? Are sources cited? Any hallucinations?
+2. **Completeness** (0-10): Does it fully address the task? Missing angles?
+3. **Structure** (0-10): Well-organized? Clear headers? Logical flow?
+4. **Actionability** (0-10): Can the reader act on this? Concrete next steps?
+5. **Conciseness** (0-10): Right level of detail? No filler?
+
+SCORING GUIDE:
+- 9-10: Exceptional. Would publish as-is.
+- 7-8: Good. Minor improvements possible.
+- 5-6: Adequate. Significant gaps or issues.
+- 3-4: Poor. Major problems.
+- 1-2: Unusable. Fundamentally flawed.
+
+RULES:
+- Be SPECIFIC. Not "could be better" but "Section 3 lacks source citations for the market size claim"
+- Be CONSTRUCTIVE. Every criticism must include a fix suggestion
+- Be HONEST. A score of 7 when it deserves 4 helps nobody
+- Evaluate the OUTPUT, not the effort
+- Consider the task type: research needs sources, code needs correctness, content needs readability
+
+OUTPUT FORMAT (EXACTLY THIS — parseable by Darwin):
+===SCORE===
+{number 1-10}
+===STRENGTHS===
+- {specific strength 1}
+- {specific strength 2}
+===WEAKNESSES===
+- {specific weakness 1 + fix suggestion}
+- {specific weakness 2 + fix suggestion}
+===VERDICT===
+{One sentence summary}
+===END===`,
+    evolution: {
+        enabled: false, // Critic doesn't evolve itself (avoids circular dependency)
+        evaluator: 'critic',
+    },
+};
+//# sourceMappingURL=critic.js.map

package/dist/agents/critic.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"critic.js","sourceRoot":"","sources":["../../src/agents/critic.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAIH,MAAM,CAAC,MAAM,MAAM,GAAoB;IACrC,IAAI,EAAE,QAAQ;IACd,IAAI,EAAE,sCAAsC;IAC5C,WAAW,EAAE,6FAA6F;IAC1G,QAAQ,EAAE,CAAC;IACX,YAAY,EAAE;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;UAqCN;IAER,SAAS,EAAE;QACT,OAAO,EAAE,KAAK,EAAG,4DAA4D;QAC7E,SAAS,EAAE,QAAQ;KACpB;CACF,CAAC"}

package/dist/agents/index.d.ts

@@ -0,0 +1,15 @@
+/**
+ * Built-in Agents — ready to use out of the box.
+ */
+export { writer } from './writer.js';
+export { researcher } from './researcher.js';
+export { critic } from './critic.js';
+export { analyst } from './analyst.js';
+export { investigator } from './investigator.js';
+export { investigatorCritic } from './investigator-critic.js';
+export { marketing } from './marketing.js';
+export { blogWriter } from './blog-writer.js';
+import type { AgentDefinition } from '../types.js';
+/** All built-in agents by name */
+export declare const builtinAgents: Record<string, AgentDefinition>;
+//# sourceMappingURL=index.d.ts.map

package/dist/agents/index.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../src/agents/index.ts"],"names":[],"mappings":"AAAA;;GAEG;AAEH,OAAO,EAAE,MAAM,EAAE,MAAM,aAAa,CAAC;AACrC,OAAO,EAAE,UAAU,EAAE,MAAM,iBAAiB,CAAC;AAC7C,OAAO,EAAE,MAAM,EAAE,MAAM,aAAa,CAAC;AACrC,OAAO,EAAE,OAAO,EAAE,MAAM,cAAc,CAAC;AACvC,OAAO,EAAE,YAAY,EAAE,MAAM,mBAAmB,CAAC;AACjD,OAAO,EAAE,kBAAkB,EAAE,MAAM,0BAA0B,CAAC;AAC9D,OAAO,EAAE,SAAS,EAAE,MAAM,gBAAgB,CAAC;AAC3C,OAAO,EAAE,UAAU,EAAE,MAAM,kBAAkB,CAAC;AAU9C,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,aAAa,CAAC;AAEnD,kCAAkC;AAClC,eAAO,MAAM,aAAa,EAAE,MAAM,CAAC,MAAM,EAAE,eAAe,CASzD,CAAC"}

package/dist/agents/index.js

@@ -0,0 +1,31 @@
+/**
+ * Built-in Agents — ready to use out of the box.
+ */
+export { writer } from './writer.js';
+export { researcher } from './researcher.js';
+export { critic } from './critic.js';
+export { analyst } from './analyst.js';
+export { investigator } from './investigator.js';
+export { investigatorCritic } from './investigator-critic.js';
+export { marketing } from './marketing.js';
+export { blogWriter } from './blog-writer.js';
+import { writer } from './writer.js';
+import { researcher } from './researcher.js';
+import { critic } from './critic.js';
+import { analyst } from './analyst.js';
+import { investigator } from './investigator.js';
+import { investigatorCritic } from './investigator-critic.js';
+import { marketing } from './marketing.js';
+import { blogWriter } from './blog-writer.js';
+/** All built-in agents by name */
+export const builtinAgents = {
+    writer,
+    researcher,
+    critic,
+    analyst,
+    investigator,
+    'investigator-critic': investigatorCritic,
+    marketing,
+    'blog-writer': blogWriter,
+};
+//# sourceMappingURL=index.js.map

package/dist/agents/index.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.js","sourceRoot":"","sources":["../../src/agents/index.ts"],"names":[],"mappings":"AAAA;;GAEG;AAEH,OAAO,EAAE,MAAM,EAAE,MAAM,aAAa,CAAC;AACrC,OAAO,EAAE,UAAU,EAAE,MAAM,iBAAiB,CAAC;AAC7C,OAAO,EAAE,MAAM,EAAE,MAAM,aAAa,CAAC;AACrC,OAAO,EAAE,OAAO,EAAE,MAAM,cAAc,CAAC;AACvC,OAAO,EAAE,YAAY,EAAE,MAAM,mBAAmB,CAAC;AACjD,OAAO,EAAE,kBAAkB,EAAE,MAAM,0BAA0B,CAAC;AAC9D,OAAO,EAAE,SAAS,EAAE,MAAM,gBAAgB,CAAC;AAC3C,OAAO,EAAE,UAAU,EAAE,MAAM,kBAAkB,CAAC;AAE9C,OAAO,EAAE,MAAM,EAAE,MAAM,aAAa,CAAC;AACrC,OAAO,EAAE,UAAU,EAAE,MAAM,iBAAiB,CAAC;AAC7C,OAAO,EAAE,MAAM,EAAE,MAAM,aAAa,CAAC;AACrC,OAAO,EAAE,OAAO,EAAE,MAAM,cAAc,CAAC;AACvC,OAAO,EAAE,YAAY,EAAE,MAAM,mBAAmB,CAAC;AACjD,OAAO,EAAE,kBAAkB,EAAE,MAAM,0BAA0B,CAAC;AAC9D,OAAO,EAAE,SAAS,EAAE,MAAM,gBAAgB,CAAC;AAC3C,OAAO,EAAE,UAAU,EAAE,MAAM,kBAAkB,CAAC;AAG9C,kCAAkC;AAClC,MAAM,CAAC,MAAM,aAAa,GAAoC;IAC5D,MAAM;IACN,UAAU;IACV,MAAM;IACN,OAAO;IACP,YAAY;IACZ,qBAAqB,EAAE,kBAAkB;IACzC,SAAS;IACT,aAAa,EAAE,UAAU;CAC1B,CAAC"}

package/dist/agents/investigator-critic.d.ts

@@ -0,0 +1,10 @@
+/**
+ * Investigator Critic — Evaluates Investigative Reports
+ *
+ * Different criteria than the standard Critic:
+ * Not "is it well written?" but
+ * "is it HONEST, BALANCED, and DEEP?"
+ */
+import type { AgentDefinition } from '../types.js';
+export declare const investigatorCritic: AgentDefinition;
+//# sourceMappingURL=investigator-critic.d.ts.map

package/dist/agents/investigator-critic.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"investigator-critic.d.ts","sourceRoot":"","sources":["../../src/agents/investigator-critic.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAEH,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,aAAa,CAAC;AAEnD,eAAO,MAAM,kBAAkB,EAAE,eAsEhC,CAAC"}

package/dist/agents/investigator-critic.js

@@ -0,0 +1,78 @@
+/**
+ * Investigator Critic — Evaluates Investigative Reports
+ *
+ * Different criteria than the standard Critic:
+ * Not "is it well written?" but
+ * "is it HONEST, BALANCED, and DEEP?"
+ */
+export const investigatorCritic = {
+    name: 'investigator-critic',
+    role: 'Investigation Quality Evaluator',
+    description: 'Evaluates investigative reports on honesty, balance, depth, and source diversity — not just writing quality.',
+    maxTurns: 5,
+    systemPrompt: `You evaluate investigative research reports on controversial topics.
+
+YOU DO NOT CARE ABOUT:
+- Pretty writing or formatting
+- Whether the conclusion is "mainstream-approved"
+- Political correctness
+- Whether it might offend someone
+
+YOU CARE ABOUT:
+1. **Honesty** (0-10): Does it say what it actually thinks, or does it hide behind diplomatic language? Does it take a position or cop out with "both sides have valid points"? Cowardly fence-sitting = low score.
+
+2. **Balance** (0-10): Are BOTH sides presented with their STRONGEST arguments? Or is one side a strawman? Does it steelman the unpopular position? Note: Balance does NOT mean 50/50 — if evidence is 80/20, a balanced report says so.
+
+3. **Source Diversity** (0-10): Does it cite ONLY mainstream sources? ONLY alternative sources? Or a genuine MIX? Are primary documents included? Are sources from different countries/perspectives? Wikipedia + CNN alone = low score. Government documents + academic papers + independent journalists = high score.
+
+4. **Depth** (0-10): Surface-level summary or actual investigation? Does it name specific people, dates, documents? Does it follow the money? Does it ask WHO BENEFITS? Generic "some experts say" = low score.
+
+5. **Intellectual Courage** (0-10): Does it challenge comfortable assumptions? Does it state uncomfortable conclusions? Does it point out what we genuinely don't know? Or does it play it safe and repeat what everyone already thinks?
+
+SCORING GUIDE:
+- 9-10: Genuinely investigative. I learned something. Both sides treated fairly. Clear honest conclusion.
+- 7-8: Good research but plays it a bit safe. Solid sources, some fence-sitting.
+- 5-6: Surface level. Mostly mainstream rehash with token mention of alternatives.
+- 3-4: Propaganda dressed as research. Only one side taken seriously.
+- 1-2: Copy-paste from Wikipedia. No actual investigation.
+
+RED FLAGS (automatic -2 penalty):
+- "This has been thoroughly debunked" without showing the debunking evidence
+- "Conspiracy theory" used as dismissal rather than description
+- No alternative sources at all
+- Conclusion was obviously predetermined before investigation
+- "More research is needed" as the ONLY conclusion
+
+GREEN FLAGS (+1 bonus):
+- Primary documents cited (leaked emails, government reports, declassified files)
+- Named specific people and their conflicts of interest
+- Changed direction mid-investigation based on evidence found
+- Stated something uncomfortable with evidence to back it up
+
+OUTPUT FORMAT (EXACTLY THIS — parseable by Darwin):
+===SCORE===
+{number 1-10}
+===HONESTY===
+{score}/10 — {specific assessment}
+===BALANCE===
+{score}/10 — {specific assessment}
+===SOURCE_DIVERSITY===
+{score}/10 — {specific assessment}
+===DEPTH===
+{score}/10 — {specific assessment}
+===COURAGE===
+{score}/10 — {specific assessment}
+===STRENGTHS===
+- {specific strength 1}
+- {specific strength 2}
+===WEAKNESSES===
+- {specific weakness 1 + what would make it better}
+- {specific weakness 2 + what would make it better}
+===VERDICT===
+{One honest sentence — was this real investigation or theatre?}
+===END===`,
+    evolution: {
+        enabled: false,
+    },
+};
+//# sourceMappingURL=investigator-critic.js.map

package/dist/agents/investigator-critic.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"investigator-critic.js","sourceRoot":"","sources":["../../src/agents/investigator-critic.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAIH,MAAM,CAAC,MAAM,kBAAkB,GAAoB;IACjD,IAAI,EAAE,qBAAqB;IAC3B,IAAI,EAAE,iCAAiC;IACvC,WAAW,EAAE,8GAA8G;IAC3H,QAAQ,EAAE,CAAC;IACX,YAAY,EAAE;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;UA4DN;IAER,SAAS,EAAE;QACT,OAAO,EAAE,KAAK;KACf;CACF,CAAC"}