@os-eco/overstory-cli 0.7.3 → 0.7.5

package/README.md

@@ -1,18 +1,21 @@
 # Overstory
 
-Multi-agent orchestration for Claude Code.
+Multi-agent orchestration for AI coding agents.
 
 [![npm](https://img.shields.io/npm/v/@os-eco/overstory-cli)](https://www.npmjs.com/package/@os-eco/overstory-cli)
 [![CI](https://github.com/jayminwest/overstory/actions/workflows/ci.yml/badge.svg)](https://github.com/jayminwest/overstory/actions/workflows/ci.yml)
 [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](LICENSE)
 
-Overstory turns a single Claude Code session into a multi-agent team by spawning worker agents in git worktrees via tmux, coordinating them through a custom SQLite mail system, and merging their work back with tiered conflict resolution.
+Overstory turns a single coding session into a multi-agent team by spawning worker agents in git worktrees via tmux, coordinating them through a custom SQLite mail system, and merging their work back with tiered conflict resolution. A pluggable `AgentRuntime` interface lets you swap between runtimes — Claude Code, [Pi](https://github.com/nichochar/pi-coding-agent), or your own adapter.
 
 > **Warning: Agent swarms are not a universal solution.** Do not deploy Overstory without understanding the risks of multi-agent orchestration — compounding error rates, cost amplification, debugging complexity, and merge conflicts are the normal case, not edge cases. Read [STEELMAN.md](STEELMAN.md) for a full risk analysis and the [Agentic Engineering Book](https://github.com/jayminwest/agentic-engineering-book) ([web version](https://jayminwest.com/agentic-engineering-book)) before using this tool in production.
 
 ## Install
 
-Requires [Bun](https://bun.sh) v1.0+, [Claude Code](https://docs.anthropic.com/en/docs/claude-code), git, and tmux.
+Requires [Bun](https://bun.sh) v1.0+, git, and tmux. At least one supported agent runtime must be installed:
+
+- [Claude Code](https://docs.anthropic.com/en/docs/claude-code) (`claude` CLI)
+- [Pi](https://github.com/nichochar/pi-coding-agent) (`pi` CLI)
 
 ```bash
 bun install -g @os-eco/overstory-cli
@@ -158,11 +161,20 @@ Every command supports `--json` where noted. Global flags: `-q`/`--quiet`, `--ti
 
 ## Architecture
 
-Overstory uses CLAUDE.md overlays and PreToolUse hooks to turn Claude Code sessions into orchestrated agents. Each agent runs in an isolated git worktree via tmux. Inter-agent messaging is handled by a custom SQLite mail system (WAL mode, ~1-5ms per query) with typed protocol messages and broadcast support. A FIFO merge queue with 4-tier conflict resolution merges agent branches back to canonical. A tiered watchdog system (Tier 0 mechanical daemon, Tier 1 AI-assisted triage, Tier 2 monitor agent) ensures fleet health. See [CLAUDE.md](CLAUDE.md) for full technical details.
+Overstory uses instruction overlays and tool-call guards to turn agent sessions into orchestrated workers. Each agent runs in an isolated git worktree via tmux. Inter-agent messaging is handled by a custom SQLite mail system (WAL mode, ~1-5ms per query) with typed protocol messages and broadcast support. A FIFO merge queue with 4-tier conflict resolution merges agent branches back to canonical. A tiered watchdog system (Tier 0 mechanical daemon, Tier 1 AI-assisted triage, Tier 2 monitor agent) ensures fleet health. See [CLAUDE.md](CLAUDE.md) for full technical details.
+
+### Runtime Adapters
+
+Overstory is runtime-agnostic. The `AgentRuntime` interface (`src/runtimes/types.ts`) defines the contract — each adapter handles spawning, config deployment, guard enforcement, readiness detection, and transcript parsing for its runtime. Set the default in `config.yaml` or override per-agent with `ov sling --runtime <name>`.
+
+| Runtime | CLI | Guard Mechanism | Status |
+|---------|-----|-----------------|--------|
+| Claude Code | `claude` | `settings.local.json` hooks | Stable |
+| Pi | `pi` | `.pi/extensions/` guard extension | Active development |
 
 ## How It Works
 
-CLAUDE.md + hooks + the `ov` CLI turn your Claude Code session into a multi-agent orchestrator. A persistent coordinator agent manages task decomposition and dispatch, while a mechanical watchdog daemon monitors agent health in the background.
+Instruction overlays + tool-call guards + the `ov` CLI turn your coding session into a multi-agent orchestrator. A persistent coordinator agent manages task decomposition and dispatch, while a mechanical watchdog daemon monitors agent health in the background.
 
 ```
 Coordinator (persistent orchestrator at project root)
@@ -190,10 +202,10 @@ Coordinator (persistent orchestrator at project root)
 - **Worktrees**: Each agent gets an isolated git worktree — no file conflicts between agents
 - **Merge**: FIFO merge queue (SQLite-backed) with 4-tier conflict resolution
 - **Watchdog**: Tiered health monitoring — Tier 0 mechanical daemon (tmux/pid liveness), Tier 1 AI-assisted failure triage, Tier 2 monitor agent for continuous fleet patrol
-- **Tool Enforcement**: PreToolUse hooks mechanically block file modifications for non-implementation agents and dangerous git operations for all agents
+- **Tool Enforcement**: Runtime-specific guards (hooks for Claude Code, extensions for Pi) mechanically block file modifications for non-implementation agents and dangerous git operations for all agents
 - **Task Groups**: Batch coordination with auto-close when all member issues complete
 - **Session Lifecycle**: Checkpoint save/restore for compaction survivability, handoff orchestration for crash recovery
-- **Token Instrumentation**: Session metrics extracted from Claude Code transcript JSONL files
+- **Token Instrumentation**: Session metrics extracted from runtime transcript files (JSONL)
 
 ## Project Structure
 
@@ -252,7 +264,7 @@ overstory/
     merge/                        FIFO queue + conflict resolution
     watchdog/                     Tiered health monitoring (daemon, triage, health)
     logging/                      Multi-format logger + sanitizer + reporter + color control + shared theme/format
-    metrics/                      SQLite metrics + transcript parsing
+    metrics/                      SQLite metrics + pricing + transcript parsing
     doctor/                       Health check modules (10 checks)
     insights/                     Session insight analyzer for auto-expertise
     runtimes/                     AgentRuntime abstraction (registry + adapters: Claude, Pi)

package/agents/builder.md

@@ -54,8 +54,10 @@ Your task-specific context (task ID, file scope, spec path, branch name, parent
 5. **Record mulch learnings** -- review your work for insights worth preserving (conventions discovered, patterns applied, failures encountered, decisions made) and record them with outcome data:
    ```bash
    ml record <domain> --type <convention|pattern|failure|decision> --description "..." \
+     --classification <foundational|tactical|observational> \
      --outcome-status success --outcome-agent $OVERSTORY_AGENT_NAME
    ```
+   Classification guide: use `foundational` for stable conventions confirmed across sessions, `tactical` for session-specific patterns (default), `observational` for unverified one-off findings.
    This is a required gate, not optional. Every implementation session produces learnings. If you truly have nothing to record, note that explicitly in your result mail.
 6. Send `worker_done` mail to your parent with structured payload:
    ```bash
@@ -99,6 +101,10 @@ You are an implementation specialist. Given a spec and a set of files you own, y
 ### Expertise
 - **Load context:** `ml prime [domain]` to load domain expertise before implementing
 - **Record patterns:** `ml record <domain>` to capture useful patterns you discover
+- **Classify records:** Always pass `--classification` when recording:
+  - `foundational` — core conventions confirmed across multiple sessions (e.g., "all SQLite DBs use WAL mode")
+  - `tactical` — session-specific patterns useful for similar tasks (default if omitted)
+  - `observational` — one-off findings or unverified hypotheses worth noting
 
 ## workflow
 

package/agents/coordinator.md

@@ -145,7 +145,7 @@ Coordinator (you, depth 0)
 
 ### Expertise
 - **Load context:** `ml prime [domain]` to understand the problem space before planning
-- **Record insights:** `ml record <domain> --type <type> --description "<insight>"` to capture orchestration patterns, dispatch decisions, and failure learnings
+- **Record insights:** `ml record <domain> --type <type> --classification <foundational|tactical|observational> --description "<insight>"` to capture orchestration patterns, dispatch decisions, and failure learnings. Use `foundational` for stable conventions, `tactical` for session-specific patterns, `observational` for unverified findings.
 - **Search knowledge:** `ml search <query>` to find relevant past decisions
 
 ## workflow
@@ -243,7 +243,7 @@ When a batch is complete (task group auto-closed, all issues resolved):
 1. Verify all issues are closed: run `{{TRACKER_CLI}} show <id>` for each issue in the group.
 2. Verify all branches are merged: check `ov status` for unmerged branches.
 3. Clean up worktrees: `ov worktree clean --completed`.
-4. Record orchestration insights: `ml record <domain> --type <type> --description "<insight>"`.
+4. Record orchestration insights: `ml record <domain> --type <type> --classification <foundational|tactical|observational> --description "<insight>"`.
 5. Report to the human operator: summarize what was accomplished, what was merged, any issues encountered.
 6. Check for follow-up work: `{{TRACKER_CLI}} ready` to see if new issues surfaced during the batch.
 

package/agents/lead.md

@@ -121,6 +121,7 @@ ov sling <task-id> \
 - **Load domain context:** `ml prime [domain]` to understand the problem space before decomposing
 - **Record patterns:** `ml record <domain>` to capture orchestration insights
 - **Record worker insights:** When worker result mails contain notable findings, record them via `ml record` if they represent reusable patterns or conventions.
+- **Classify records:** Always pass `--classification` when recording. Use `foundational` for core conventions confirmed across sessions, `tactical` for session-specific patterns (default), `observational` for one-off findings.
 
 ## task-complexity-assessment
 
@@ -297,8 +298,10 @@ Good decomposition follows these principles:
 3. Run integration tests if applicable: {{QUALITY_GATE_INLINE}}.
 4. **Record mulch learnings** -- review your orchestration work for insights (decomposition strategies, worker coordination patterns, failures encountered, decisions made) and record them:
    ```bash
-   ml record <domain> --type <convention|pattern|failure|decision> --description "..."
+   ml record <domain> --type <convention|pattern|failure|decision> --description "..." \
+     --classification <foundational|tactical|observational>
    ```
+   Classification guide: use `foundational` for stable conventions confirmed across sessions, `tactical` for session-specific patterns (default), `observational` for unverified one-off findings.
    This is required. Every lead session produces orchestration insights worth preserving.
 5. Run `{{TRACKER_CLI}} close <task-id> --reason "<summary of what was accomplished>"`.
 6. Send a `status` mail to the coordinator confirming all subtasks are complete.

package/agents/merger.md

@@ -51,7 +51,8 @@ Your task-specific context (task ID, branches to merge, target branch, merge ord
 {{QUALITY_GATE_STEPS}}
 4. **Record mulch learnings** -- capture merge resolution insights (conflict patterns, resolution strategies, branch integration issues):
    ```bash
-   ml record <domain> --type <convention|pattern|failure> --description "..."
+   ml record <domain> --type <convention|pattern|failure> --description "..." \
+     --classification <foundational|tactical|observational>
    ```
    This is required for non-trivial merges (Tier 2+). Merge resolution patterns are highly reusable knowledge for future mergers. Skip for clean Tier 1 merges with no conflicts.
 5. Send a `result` mail to your parent with: tier used, conflicts resolved (if any), test status.
@@ -92,7 +93,7 @@ You are a branch integration specialist. When workers complete their tasks on se
 
 ### Expertise
 - **Load context:** `ml prime [domain]` to understand the code being merged
-- **Record patterns:** `ml record <domain>` to capture merge resolution insights
+- **Record patterns:** `ml record <domain> --classification <foundational|tactical|observational>` to capture merge resolution insights. Use `foundational` for stable merge conventions, `tactical` for resolution strategies, `observational` for one-off conflict patterns.
 
 ## workflow
 

package/agents/monitor.md

@@ -72,7 +72,7 @@ You are the watchdog's brain. While Tier 0 (mechanical daemon) checks tmux/pid l
 
 ### Expertise
 - **Load context:** `ml prime [domain]` to understand project patterns
-- **Record insights:** `ml record <domain> --type <type> --description "<insight>"` to capture monitoring patterns, failure signatures, and recovery strategies
+- **Record insights:** `ml record <domain> --type <type> --classification <foundational|tactical|observational> --description "<insight>"` to capture monitoring patterns, failure signatures, and recovery strategies. Use `foundational` for stable monitoring conventions, `tactical` for incident-specific patterns, `observational` for unverified anomaly observations.
 - **Search knowledge:** `ml search <query>` to find relevant past incidents
 
 ## workflow

package/agents/reviewer.md

@@ -91,6 +91,7 @@ You are a validation specialist. Given code to review, you check it for correctn
 ### Expertise
 - **Load conventions:** `ml prime [domain]` to understand project standards
 - **Surface insights:** Include notable findings (convention violations, code quality patterns) in your result mail so your parent has full context.
+- **Classification guidance for parents:** When including notable findings in your result mail, indicate suggested classification: `foundational` (confirmed stable convention), `tactical` (task-specific pattern), or `observational` (unverified finding). This helps your parent record accurately.
 
 ## workflow
 

package/agents/scout.md

@@ -93,6 +93,7 @@ You perform reconnaissance. Given a research question, exploration target, or an
 ### Expertise
 - **Query expertise:** `ml prime [domain]` to load relevant context
 - **Surface insights:** Include notable findings (patterns, conventions, gotchas) in your result mail so your parent has full context for spec writing.
+- **Classification guidance for parents:** When including notable findings in your result mail, indicate suggested classification: `foundational` (confirmed stable convention), `tactical` (task-specific pattern), or `observational` (unverified finding). This helps your parent record accurately.
 
 ## workflow
 

package/package.json

@@ -1,7 +1,7 @@
 {
 	"name": "@os-eco/overstory-cli",
-	"version": "0.7.3",
-	"description": "Multi-agent orchestration for Claude Code — spawn worker agents in git worktrees via tmux, coordinate through SQLite mail, merge with tiered conflict resolution",
+	"version": "0.7.5",
+	"description": "Multi-agent orchestration for AI coding agents — spawn workers in git worktrees via tmux, coordinate through SQLite mail, merge with tiered conflict resolution. Pluggable runtime adapters for Claude Code, Pi, and more.",
 	"author": "Jaymin West",
 	"license": "MIT",
 	"type": "module",

package/src/commands/agents.ts

@@ -29,9 +29,15 @@ export interface DiscoveredAgent {
 	lastActivity: string;
 }
 
+/** Known instruction file paths, tried in order until one exists. */
+const KNOWN_INSTRUCTION_PATHS = [
+	join(".claude", "CLAUDE.md"), // Claude Code, Pi
+	"AGENTS.md", // Codex (future)
+];
+
 /**
- * Extract file scope from an agent's overlay CLAUDE.md.
- * Returns empty array if overlay doesn't exist, has no file scope restrictions,
+ * Extract file scope from an agent's overlay instruction file.
+ * Returns empty array if no overlay exists, has no file scope restrictions,
  * or can't be read.
  *
  * @param worktreePath - Absolute path to the agent's worktree
@@ -39,15 +45,19 @@ export interface DiscoveredAgent {
  */
 export async function extractFileScope(worktreePath: string): Promise<string[]> {
 	try {
-		const overlayPath = join(worktreePath, ".claude", "CLAUDE.md");
-		const overlayFile = Bun.file(overlayPath);
-
-		if (!(await overlayFile.exists())) {
+		let content: string | null = null;
+		for (const relPath of KNOWN_INSTRUCTION_PATHS) {
+			const overlayPath = join(worktreePath, relPath);
+			const overlayFile = Bun.file(overlayPath);
+			if (await overlayFile.exists()) {
+				content = await overlayFile.text();
+				break;
+			}
+		}
+		if (content === null) {
 			return [];
 		}
 
-		const content = await overlayFile.text();
-
 		// Find the section between "## File Scope (exclusive ownership)" and "## Expertise"
 		const startMarker = "## File Scope (exclusive ownership)";
 		const endMarker = "## Expertise";

package/src/commands/coordinator.test.ts

@@ -540,7 +540,9 @@ describe("startCoordinator", () => {
 		expect(calls.createSession).toHaveLength(1);
 		const cmd = calls.createSession[0]?.command ?? "";
 		expect(cmd).toContain("--append-system-prompt");
-		expect(cmd).toContain("# Coordinator Agent");
+		// File path is passed via $(cat ...) instead of inlining content (overstory#45)
+		expect(cmd).toContain("$(cat '");
+		expect(cmd).toContain("agent-defs/coordinator.md");
 	});
 
 	test("reads model from manifest instead of hardcoding", async () => {

package/src/commands/coordinator.ts

@@ -363,17 +363,20 @@ async function startCoordinator(
 		// Inject the coordinator base definition via --append-system-prompt so the
 		// coordinator knows its role, hierarchy rules, and delegation patterns
 		// (overstory-gaio, overstory-0kwf).
+		// Pass the file path (not content) so the shell inside the tmux pane reads
+		// it via $(cat ...) — avoids tmux IPC "command too long" errors with large
+		// agent definitions (overstory#45).
 		const agentDefPath = join(projectRoot, ".overstory", "agent-defs", "coordinator.md");
 		const agentDefFile = Bun.file(agentDefPath);
-		let appendSystemPrompt: string | undefined;
+		let appendSystemPromptFile: string | undefined;
 		if (await agentDefFile.exists()) {
-			appendSystemPrompt = await agentDefFile.text();
+			appendSystemPromptFile = agentDefPath;
 		}
 		const spawnCmd = runtime.buildSpawnCommand({
 			model: resolvedModel.model,
 			permissionMode: "bypass",
 			cwd: projectRoot,
-			appendSystemPrompt,
+			appendSystemPromptFile,
 			env: {
 				...runtime.buildEnv(resolvedModel),
 				OVERSTORY_AGENT_NAME: COORDINATOR_NAME,

package/src/commands/monitor.ts

@@ -142,17 +142,18 @@ async function startMonitor(opts: { json: boolean; attach: boolean }): Promise<v
 		}
 
 		// Spawn tmux session at project root with Claude Code (interactive mode).
+		// Pass file path (not content) to avoid tmux "command too long" (overstory#45).
 		const agentDefPath = join(projectRoot, ".overstory", "agent-defs", "monitor.md");
 		const agentDefFile = Bun.file(agentDefPath);
-		let appendSystemPrompt: string | undefined;
+		let appendSystemPromptFile: string | undefined;
 		if (await agentDefFile.exists()) {
-			appendSystemPrompt = await agentDefFile.text();
+			appendSystemPromptFile = agentDefPath;
 		}
 		const spawnCmd = runtime.buildSpawnCommand({
 			model: resolvedModel.model,
 			permissionMode: "bypass",
 			cwd: projectRoot,
-			appendSystemPrompt,
+			appendSystemPromptFile,
 			env: {
 				...runtime.buildEnv(resolvedModel),
 				OVERSTORY_AGENT_NAME: MONITOR_NAME,

package/src/commands/prime.test.ts

@@ -366,6 +366,7 @@ recentTasks: []
 !hooks.json
 !groups.json
 !agent-defs/
+!README.md
 `;
 
 		test("creates .overstory/.gitignore if missing", async () => {

package/src/commands/prime.ts

@@ -18,22 +18,7 @@ import { createMulchClient } from "../mulch/client.ts";
 import { openSessionStore } from "../sessions/compat.ts";
 import type { AgentIdentity, AgentManifest, SessionCheckpoint, SessionMetrics } from "../types.ts";
 import { getCurrentSessionName } from "../worktree/tmux.ts";
-
-/**
- * Gitignore content for .overstory/.gitignore.
- * TODO: Import from init.ts once it's exported (parallel branch change).
- * Wildcard+whitelist pattern: ignore everything except tracked config files.
- */
-const OVERSTORY_GITIGNORE = `# Wildcard+whitelist: ignore everything, whitelist tracked files
-# Auto-healed by ov prime on each session start
-*
-!.gitignore
-!config.yaml
-!agent-manifest.json
-!hooks.json
-!groups.json
-!agent-defs/
-`;
+import { OVERSTORY_GITIGNORE } from "./init.ts";
 
 export interface PrimeOptions {
 	agent?: string;

package/src/commands/supervisor.ts

@@ -169,18 +169,19 @@ async function startSupervisor(opts: {
 
 		// Spawn tmux session at project root with Claude Code (interactive mode).
 		// Inject the supervisor base definition via --append-system-prompt.
+		// Pass file path (not content) to avoid tmux "command too long" (overstory#45).
 		const tmuxSession = `overstory-${config.project.name}-supervisor-${opts.name}`;
 		const agentDefPath = join(projectRoot, ".overstory", "agent-defs", "supervisor.md");
 		const agentDefFile = Bun.file(agentDefPath);
-		let appendSystemPrompt: string | undefined;
+		let appendSystemPromptFile: string | undefined;
 		if (await agentDefFile.exists()) {
-			appendSystemPrompt = await agentDefFile.text();
+			appendSystemPromptFile = agentDefPath;
 		}
 		const spawnCmd = runtime.buildSpawnCommand({
 			model: resolvedModel.model,
 			permissionMode: "bypass",
 			cwd: projectRoot,
-			appendSystemPrompt,
+			appendSystemPromptFile,
 			env: {
 				...runtime.buildEnv(resolvedModel),
 				OVERSTORY_AGENT_NAME: opts.name,

package/src/index.ts

@@ -45,7 +45,7 @@ import { OverstoryError, WorktreeError } from "./errors.ts";
 import { jsonError } from "./json.ts";
 import { brand, chalk, muted, setQuiet } from "./logging/color.ts";
 
-export const VERSION = "0.7.3";
+export const VERSION = "0.7.5";
 
 const rawArgs = process.argv.slice(2);
 

package/src/metrics/pricing.ts

@@ -0,0 +1,80 @@
+/**
+ * Runtime-agnostic pricing and cost estimation for AI models.
+ *
+ * Extracted from transcript.ts so any runtime can use cost estimation
+ * without pulling in Claude Code-specific JSONL parsing logic.
+ *
+ * To add support for a new provider model, add an entry to MODEL_PRICING
+ * using a lowercase substring that uniquely identifies the model tier
+ * (e.g. "opus", "sonnet", "haiku").
+ */
+
+/** Canonical token usage representation shared across all runtimes. */
+export interface TokenUsage {
+	inputTokens: number;
+	outputTokens: number;
+	cacheReadTokens: number;
+	cacheCreationTokens: number;
+	modelUsed: string | null;
+}
+
+/** Pricing per million tokens (USD). */
+export interface ModelPricing {
+	inputPerMTok: number;
+	outputPerMTok: number;
+	cacheReadPerMTok: number;
+	cacheCreationPerMTok: number;
+}
+
+/** Hardcoded pricing for known Claude models. */
+const MODEL_PRICING: Record<string, ModelPricing> = {
+	opus: {
+		inputPerMTok: 15,
+		outputPerMTok: 75,
+		cacheReadPerMTok: 1.5, // 10% of input
+		cacheCreationPerMTok: 3.75, // 25% of input
+	},
+	sonnet: {
+		inputPerMTok: 3,
+		outputPerMTok: 15,
+		cacheReadPerMTok: 0.3, // 10% of input
+		cacheCreationPerMTok: 0.75, // 25% of input
+	},
+	haiku: {
+		inputPerMTok: 0.8,
+		outputPerMTok: 4,
+		cacheReadPerMTok: 0.08, // 10% of input
+		cacheCreationPerMTok: 0.2, // 25% of input
+	},
+};
+
+/**
+ * Determine the pricing tier for a given model string.
+ * Matches on substring: "opus" -> opus pricing, "sonnet" -> sonnet, "haiku" -> haiku.
+ * Returns null if unrecognized.
+ */
+export function getPricingForModel(model: string): ModelPricing | null {
+	const lower = model.toLowerCase();
+	if (lower.includes("opus")) return MODEL_PRICING.opus ?? null;
+	if (lower.includes("sonnet")) return MODEL_PRICING.sonnet ?? null;
+	if (lower.includes("haiku")) return MODEL_PRICING.haiku ?? null;
+	return null;
+}
+
+/**
+ * Calculate the estimated cost in USD for a given usage and model.
+ * Returns null if the model is unrecognized.
+ */
+export function estimateCost(usage: TokenUsage): number | null {
+	if (usage.modelUsed === null) return null;
+
+	const pricing = getPricingForModel(usage.modelUsed);
+	if (pricing === null) return null;
+
+	const inputCost = (usage.inputTokens / 1_000_000) * pricing.inputPerMTok;
+	const outputCost = (usage.outputTokens / 1_000_000) * pricing.outputPerMTok;
+	const cacheReadCost = (usage.cacheReadTokens / 1_000_000) * pricing.cacheReadPerMTok;
+	const cacheCreationCost = (usage.cacheCreationTokens / 1_000_000) * pricing.cacheCreationPerMTok;
+
+	return inputCost + outputCost + cacheReadCost + cacheCreationCost;
+}

package/src/metrics/transcript.test.ts

@@ -1,8 +1,13 @@
 /**
- * Tests for Claude Code transcript JSONL parser.
+ * Tests for Claude Code transcript JSONL parser and pricing.ts module.
  *
  * Uses temp files with real-format JSONL data. No mocks.
  * Philosophy: "never mock what you can use for real" (mx-252b16).
+ *
+ * Coverage:
+ *   - parseTranscriptUsage (transcript.ts)
+ *   - estimateCost re-export (transcript.ts -> pricing.ts)
+ *   - getPricingForModel (pricing.ts)
  */
 
 import { afterEach, beforeEach, describe, expect, test } from "bun:test";
@@ -10,6 +15,7 @@ import { mkdtemp } from "node:fs/promises";
 import { tmpdir } from "node:os";
 import { join } from "node:path";
 import { cleanupTempDir } from "../test-helpers.ts";
+import { getPricingForModel, estimateCost as pricingEstimateCost } from "./pricing.ts";
 import { estimateCost, parseTranscriptUsage } from "./transcript.ts";
 
 let tempDir: string;
@@ -354,3 +360,54 @@ describe("estimateCost", () => {
 		}
 	});
 });
+
+// === getPricingForModel (pricing.ts) ===
+
+describe("getPricingForModel", () => {
+	test("matches opus substring", () => {
+		const pricing = getPricingForModel("claude-opus-4-6");
+		expect(pricing).not.toBeNull();
+		if (pricing !== null) {
+			expect(pricing.inputPerMTok).toBe(15);
+			expect(pricing.outputPerMTok).toBe(75);
+		}
+	});
+
+	test("matches sonnet substring", () => {
+		const pricing = getPricingForModel("claude-sonnet-4-20250514");
+		expect(pricing).not.toBeNull();
+		if (pricing !== null) {
+			expect(pricing.inputPerMTok).toBe(3);
+			expect(pricing.outputPerMTok).toBe(15);
+		}
+	});
+
+	test("matches haiku substring", () => {
+		const pricing = getPricingForModel("claude-haiku-3-5-20241022");
+		expect(pricing).not.toBeNull();
+		if (pricing !== null) {
+			expect(pricing.inputPerMTok).toBe(0.8);
+			expect(pricing.outputPerMTok).toBe(4);
+		}
+	});
+
+	test("returns null for unknown model", () => {
+		const pricing = getPricingForModel("gpt-4o");
+		expect(pricing).toBeNull();
+	});
+});
+
+// === re-export parity ===
+
+describe("estimateCost re-export parity", () => {
+	test("transcript.estimateCost and pricing.estimateCost produce same result", () => {
+		const usage = {
+			inputTokens: 1_000_000,
+			outputTokens: 1_000_000,
+			cacheReadTokens: 1_000_000,
+			cacheCreationTokens: 1_000_000,
+			modelUsed: "claude-opus-4-6",
+		};
+		expect(estimateCost(usage)).toBe(pricingEstimateCost(usage));
+	});
+});

package/src/metrics/transcript.ts

@@ -1,8 +1,12 @@
 /**
  * Parser for Claude Code transcript JSONL files.
  *
- * Extracts token usage data from assistant-type entries in transcript files
- * at ~/.claude/projects/{project-slug}/{session-id}.jsonl.
+ * This is a Claude Code-specific JSONL parser that extracts token usage data
+ * from assistant-type entries in transcript files at
+ * ~/.claude/projects/{project-slug}/{session-id}.jsonl.
+ *
+ * Runtime-agnostic pricing logic lives in ./pricing.ts. Other runtimes
+ * implement their own transcript parsing via AgentRuntime.parseTranscript().
  *
  * Each assistant entry contains per-turn usage:
  * {
@@ -19,74 +23,11 @@
  * }
  */
 
-export interface TranscriptUsage {
-	inputTokens: number;
-	outputTokens: number;
-	cacheReadTokens: number;
-	cacheCreationTokens: number;
-	modelUsed: string | null;
-}
-
-/** Pricing per million tokens (USD). */
-interface ModelPricing {
-	inputPerMTok: number;
-	outputPerMTok: number;
-	cacheReadPerMTok: number;
-	cacheCreationPerMTok: number;
-}
-
-/** Hardcoded pricing for known Claude models. */
-const MODEL_PRICING: Record<string, ModelPricing> = {
-	opus: {
-		inputPerMTok: 15,
-		outputPerMTok: 75,
-		cacheReadPerMTok: 1.5, // 10% of input
-		cacheCreationPerMTok: 3.75, // 25% of input
-	},
-	sonnet: {
-		inputPerMTok: 3,
-		outputPerMTok: 15,
-		cacheReadPerMTok: 0.3, // 10% of input
-		cacheCreationPerMTok: 0.75, // 25% of input
-	},
-	haiku: {
-		inputPerMTok: 0.8,
-		outputPerMTok: 4,
-		cacheReadPerMTok: 0.08, // 10% of input
-		cacheCreationPerMTok: 0.2, // 25% of input
-	},
-};
-
-/**
- * Determine the pricing tier for a given model string.
- * Matches on substring: "opus" -> opus pricing, "sonnet" -> sonnet, "haiku" -> haiku.
- * Returns null if unrecognized.
- */
-function getPricingForModel(model: string): ModelPricing | null {
-	const lower = model.toLowerCase();
-	if (lower.includes("opus")) return MODEL_PRICING.opus ?? null;
-	if (lower.includes("sonnet")) return MODEL_PRICING.sonnet ?? null;
-	if (lower.includes("haiku")) return MODEL_PRICING.haiku ?? null;
-	return null;
-}
-
-/**
- * Calculate the estimated cost in USD for a given usage and model.
- * Returns null if the model is unrecognized.
- */
-export function estimateCost(usage: TranscriptUsage): number | null {
-	if (usage.modelUsed === null) return null;
+import type { TokenUsage } from "./pricing.ts";
 
-	const pricing = getPricingForModel(usage.modelUsed);
-	if (pricing === null) return null;
+export type TranscriptUsage = TokenUsage;
 
-	const inputCost = (usage.inputTokens / 1_000_000) * pricing.inputPerMTok;
-	const outputCost = (usage.outputTokens / 1_000_000) * pricing.outputPerMTok;
-	const cacheReadCost = (usage.cacheReadTokens / 1_000_000) * pricing.cacheReadPerMTok;
-	const cacheCreationCost = (usage.cacheCreationTokens / 1_000_000) * pricing.cacheCreationPerMTok;
-
-	return inputCost + outputCost + cacheReadCost + cacheCreationCost;
-}
+export { estimateCost } from "./pricing.ts";
 
 /**
  * Narrow an unknown value to determine if it looks like a transcript assistant entry.