Changelog

Autonomously-generated entries from agent runs. Most recent first.

feat(munchkins): launch agent runs inside cmux workspace when available (86ddc5a)

2026-05-10 19:35 PDT · feat-small · 604.6s · $4.0926

Goal: When cmux is on PATH, bun run munchkins <agent> ... should re-launch the same invocation inside a fresh cmux new-workspace and exit immediately; otherwise behavior is unchanged.

Outcome: Added packages/munchkins/src/cmux-launcher.ts exporting two pure helpers (shouldDelegateToCmux, buildCmuxCommand) plus a Bun-test suite covering delegation gating and POSIX-safe command construction. Wired the pre-check into packages/munchkins/src/index.ts ahead of the existing registry dispatch, strips --no-cmux from argv before commander sees it, and extended the scenario harness audit guard to also reject real cmux invocations.

How to test manually:

From the repo root run bun run lint && bun run typecheck && bun test packages/munchkins/src — all three should pass; the new cmux-launcher.test.ts suite must be green.
Run bun run scenario — should pass; the audit guard now also bans cmux invocations, but no scenario invokes one so there is no regression.
With cmux NOT on PATH (verify via which cmux → empty), run bun run munchkins bug-fix --user-message=./scratch/example.md (create a tiny scratch file first). It should execute inline exactly as before — same stdout, same worktree behavior.
With cmux installed and the cmux app running, run the same bun run munchkins bug-fix --user-message=./scratch/example.md. Expect a single stdout line of the form Launching bug-fix in cmux workspace: bug-fix-<timestamp> and the outer shell to return promptly. Open the cmux app and confirm a new workspace named bug-fix-<timestamp> exists with the agent running inside it.
Opt-out via env: with cmux installed, run MUNCHKINS_NO_CMUX=1 bun run munchkins bug-fix --user-message=./scratch/example.md. It should run inline (no Launching ... line, no new workspace).
Opt-out via flag: with cmux installed, run bun run munchkins bug-fix --user-message=./scratch/example.md --no-cmux. It should run inline, and commander must NOT error on the unknown flag (the flag is stripped before parsing). Also verify bun run munchkins bug-fix --help does not list --no-cmux.
Meta/help paths stay inline with cmux installed — verify each prints normally without opening a workspace: bun run munchkins --help, bun run munchkins, bun run munchkins daemon (cancel with Ctrl-C), bun run munchkins resume, bun run munchkins status, bun run munchkins skills install, bun run munchkins bug-fix --dry-run, bun run munchkins bug-fix --help.
Edge case for shell escaping: with cmux installed, run bun run munchkins bug-fix --user-message="can't stop" and confirm the workspace launches successfully (the single quote in the value is POSIX-escaped via '\'' inside the --command payload, covered by cmux-launcher.test.ts as well).

Files changed:

packages/munchkins/package.json
packages/munchkins/src/cmux-launcher.ts
packages/munchkins/src/cmux-launcher.test.ts
packages/munchkins/src/index.ts
scenarios/lib/mock-spawn-claude.ts

feat(munchkins-core): add Prompt.withSkill() helper (07441da)

2026-05-10 19:34 PDT · feat-small · 385.0s · $2.3356

Goal: Add a Prompt.withSkill(name) helper that reads <repoRoot>/.claude/skills/<name>/SKILL.md, strips YAML frontmatter, and contributes the body to the system prompt — same slot as withSystem(path).

Outcome: Refactored Prompt's internal systemPaths: string[] into a tagged systemSources array supporting both path and skill source kinds. withSystem(path) semantics are unchanged; the new withSkill(name) queues a skill source that resolves to .claude/skills/<name>/SKILL.md at resolve() time, with a textual frontmatter strip (including one trailing blank line). Missing-skill and malformed-frontmatter errors carry actionable messages pointing at bun run munchkins install-skills and the offending path. Eight test cases in the new prompt.test.ts cover stripping, chaining, errors, composition with withSystem, and preservation of pre-change withSystem behavior.

How to test manually:

From the repo root, run the new test file directly: bun test packages/munchkins-core/src/builder/prompt.test.ts — expect all 8 cases to pass.
Run the broader gate: bun run lint && bun run typecheck && bun run test — expect green; existing tests must still pass since withSystem(path) behavior is preserved.

Verify the skill happy path against a real on-disk file. Create a fixture and a one-liner script:

mkdir -p /tmp/wskill-demo/.claude/skills/demo
printf -- '---\nname: demo\ndescription: x\n---\n\n# Demo body\n' > /tmp/wskill-demo/.claude/skills/demo/SKILL.md
bun -e 'import {Prompt} from "./packages/munchkins-core/src/builder/prompt.ts"; console.log(JSON.stringify(new Prompt().withSkill("demo").resolve("/tmp/wskill-demo")))'

Expected: systemPrompt equals "# Demo body\n" (no frontmatter, no leading blank line).

Edge: missing skill. Run the same one-liner but with withSkill("missing") — expect a thrown error whose message contains Skill 'missing' not found and install-skills.
Edge: malformed frontmatter. Overwrite the file with printf -- '---\nname: demo\nno close here' > /tmp/wskill-demo/.claude/skills/demo/SKILL.md and re-run the one-liner — expect an error containing malformed frontmatter (no closing '---' delimiter) and the absolute path to the file.
Edge: composition order. Create /tmp/wskill-demo/a.md (AAA) and /tmp/wskill-demo/b.md (BBB), then chain new Prompt().withSystem("/tmp/wskill-demo/a.md").withSkill("demo").withSystem("/tmp/wskill-demo/b.md").resolve("/tmp/wskill-demo") — expect systemPrompt to be AAA\n\n# Demo body\n\n\nBBB (sources joined with \n\n in call order).
Sanity: confirm existing agents still work without changes — bun run scenario should pass, exercising withSystem(path) callers in bugfix-agent.ts, refactor-agent.ts, feat-small-agent.ts, and bugfix-then-refactor-agent.ts.

Files changed:

packages/munchkins-core/src/builder/prompt.ts
packages/munchkins-core/src/builder/prompt.test.ts

feat(munchkins): add SKILL.md discovery surface for default agents (bd12fe3)

2026-05-10 19:20 PDT · feat-small · 265.9s · $2.0450

Goal: Add Claude Code SKILL.md files for the three default munchkins agents (bug-fix, refactor, feat-small) and symlink them into .claude/skills/ so Claude Code discovers them as /<name>.

Outcome: Created three new SKILL.md files under packages/munchkins/skills/<name>/ with YAML frontmatter (name, description) followed by the body copied byte-for-byte from each agent's source prompt md. Added three relative symlinks under .claude/skills/ pointing at the corresponding skill directories. No agent .ts files were modified — content is duplicated with agents/<name>/prompts/<name>.md for this MVP; the later migration will collapse the duplication.

How to test manually:

From the repo root, confirm the three new SKILL.md files exist:

ls packages/munchkins/skills/bug-fix/SKILL.md packages/munchkins/skills/refactor/SKILL.md packages/munchkins/skills/feat-small/SKILL.md

Expected: all three paths print with no error.

Verify the frontmatter is well-formed and only contains name and description:
```
head -5 packages/munchkins/skills/bug-fix/SKILL.md
head -5 packages/munchkins/skills/refactor/SKILL.md
head -5 packages/munchkins/skills/feat-small/SKILL.md
```
Expected: each starts with ---, has name: <slug> matching the directory, description: ... (one sentence), and closes with ---.

Verify the body of each SKILL.md is byte-identical to the source prompt md (skipping the 4-line frontmatter + blank line):

diff <(tail -n +6 packages/munchkins/skills/bug-fix/SKILL.md) packages/munchkins/agents/bugfix/prompts/bug-fix.md
diff <(tail -n +6 packages/munchkins/skills/refactor/SKILL.md) packages/munchkins/agents/refactor/prompts/refactor.md
diff <(tail -n +6 packages/munchkins/skills/feat-small/SKILL.md) packages/munchkins/agents/feat-small/prompts/feat-small.md

Expected: all three diffs produce no output.

Verify each symlink resolves to the expected relative target:
```
readlink .claude/skills/bug-fix
readlink .claude/skills/refactor
readlink .claude/skills/feat-small
```
Expected: each prints ../../packages/munchkins/skills/<name>.
Verify the existing agents still register and run:
```
bun run munchkins --help
```
Expected: usage lists bug-fix, refactor, and feat-small subcommands (no regression).

Verify the source prompt files were not modified by the change:

git diff HEAD~1 -- packages/munchkins/agents/bugfix/prompts/bug-fix.md packages/munchkins/agents/refactor/prompts/refactor.md packages/munchkins/agents/feat-small/prompts/feat-small.md

Expected: no output (files unchanged in this commit).

Out-of-band Claude Code discovery check: open this repo in Claude Code and confirm that typing /bug-fix, /refactor, and /feat-small each surface the new skills with their full descriptions (matches the strings in packages/munchkins/skills/<name>/SKILL.md).

Files changed:

packages/munchkins/skills/bug-fix/SKILL.md
packages/munchkins/skills/refactor/SKILL.md
packages/munchkins/skills/feat-small/SKILL.md
.claude/skills/bug-fix (symlink)
.claude/skills/refactor (symlink)
.claude/skills/feat-small (symlink)

feat(docs): re-orient onboarding around the new-munchkin skill (deb5605)

2026-05-10 14:10 PDT · feat-small · 458.8s · $3.2744

Goal: Re-orient the docs onboarding so the new-munchkin skill is the destination, with the bug-fix run demoted to a proof-of-life smoke test.

Outcome: Restructured docs/pages/getting-started.md into six sections (Prerequisites, Install, Proof of life, Scaffold your first agent, Next steps), added a fifth Claude Code prerequisite, surfaced bun run munchkins skills install, and compressed the artifact tree and failure recovery to one-liners. Updated docs/pages/index.mdx to drop the working-guide line, split the CTA row into Defaults vs. Build-your-own, and rewrite the Get-started gloss. Reordered docs/pages/agents/custom.md so the new-munchkin skill section leads (now 3 paragraphs covering trigger phrases, repo introspection, and create-mode outputs) and the manual path follows. Reordered docs/pages/agents/_meta.json so custom is first.

How to test manually:

From the repo root, run bun run docs:dev and open the docs site in a browser.
Land on / and confirm the lede now reads "The default agents are working examples…"; verify the CTA row shows three lines: Get started, Defaults (reference): …, and Build your own: /new-munchkin skill · AgentBuilder API. The headline and proof-tail should be unchanged.
Click Get started. Confirm the page has six top-level ## headings in this exact order: Prerequisites, Install, Proof of life: run the bug-fix agent, Scaffold your first agent for this repo, Next steps. (Plus the # Getting started title.) The old "Where the artifacts go" tree and "If it fails" subsection should be gone.
In Prerequisites, confirm there are five items and the fifth names Claude Code as optional.
In Install, confirm "munchkins": "munchkins" is the only script line shown, the bunx alternative is mentioned in one sentence, and bun run munchkins skills install appears exactly once.
In the new "Scaffold your first agent" section, confirm /new-munchkin (with the slash) is shown in a fenced block and the three create-mode bullets are present.
Click into Agents from the sidebar. Confirm the order is Build your own, Bug fix, Small feature, Refactor.
Open Build your own. Confirm the first ## after the title is "Scaffold with the new-munchkin skill" and contains 3 paragraphs; the second ## is "What you're building" and starts with the bridging sentence about the manual path.
Edge case: grep the rendered docs for the deleted line This site is a working guide to the framework — it should return zero hits. Also grep getting-started.md to confirm /new-munchkin appears at least once and bun run munchkins skills install appears exactly once.
Run bun run munchkins skills install --dest /tmp/munchkins-skills-test and confirm the bundled skills land at the override path — this validates that the new Install step works as documented.

Files changed:

docs/pages/getting-started.md
docs/pages/index.mdx
docs/pages/agents/custom.md
docs/pages/agents/_meta.json

docs(pages): add user-facing agent guide organized by agent (a01c8fb)

2026-05-10 13:31 PDT · feat-small · 445.7s · $7.0202

Goal: Replace the thin Rspress site with a full user-facing guide where each default agent has its own self-contained page, plus getting-started and custom-agent pages.

Outcome: Added getting-started.md, agents/{bug-fix,feat-small,refactor,custom}.md, and agents/_meta.json; rewrote index.mdx as a real landing page; updated root _meta.json to surface the new sections. The three default-agent pages each follow the 14-section skeleton (substantive content, intentional repetition for self-containment); custom.md covers every public method on AgentBuilder and Prompt. Also bumped two agent-builder.test.ts integration tests to a 30s timeout so the real-git E2E cases don't graze the default 5s limit under parallel load.

How to test manually:

From the repo root, run PUBLIC_DOCS=true bun run docs:build and confirm it exits 0. Verify the four agent pages and getting-started are emitted under docs/doc_build/ (or wherever Rspress writes output) and that internal/** is excluded.
Run bun --cwd docs run dev and open the site in a browser. Confirm the sidebar order: Home, Getting started, Agents (expandable to Bug fix, Small feature, Refactor, Build your own), Changelog. The Internal section should still render in dev mode.
From the home page, click Get started — expect to land on /getting-started. Click Bug fix — expect /agents/bug-fix. Each default-agent page should scroll through all 14 sections in order (What it does → Worked example).
On /agents/bug-fix, ctrl-F for --user-message, --cli, --integrate, --dry-run, --thinking, --verbose — all six must appear in the Flags table. Repeat on feat-small.md and refactor.md.
On /agents/custom, ctrl-F for each AgentBuilder method (option, add, addDeterministic, summaryWriter, integrate, setSandbox, rename, describe, thenRun, cron, run, runFromState) and each Prompt method (withSystem, withUserMessage, withUserMessageFromOption). All must be present. Confirm launch-munchkin and new-munchkin are both referenced.
Search across all four new pages for MUNCHKINS_CLI, MUNCHKINS_RUN_LOG_DIR, and MUNCHKINS_CHANGELOG_PATH — each env var should appear at least once across the set.
Confirm bun run munchkins resume --list | --latest | <id> syntax appears in each of the three default-agent pages, and bun run munchkins daemon + .cron() appear in all three default-agent pages and in custom.md. bun run munchkins skills install [--dest <path>] should appear in custom.md.
Run bun run lint, bun run typecheck, and bun run scenario from the repo root — all should pass. (The deterministic gate in CI does this automatically.)
Edge case: open docs/pages/changelog.md and docs/pages/internal/** in git — confirm they are unchanged by this commit (git diff HEAD~1 -- docs/pages/changelog.md docs/pages/internal/).
Edge case: run bun test packages/munchkins-core/src/builder/agent-builder.test.ts and confirm the two integration tests in AgentBuilder.run integration dispatch end-to-end pass without timeout warnings.

Files changed:

docs/pages/_meta.json
docs/pages/index.mdx
docs/pages/getting-started.md
docs/pages/agents/_meta.json
docs/pages/agents/bug-fix.md
docs/pages/agents/feat-small.md
docs/pages/agents/refactor.md
docs/pages/agents/custom.md

fix(munchkins): make package bunx-executable via bin entry (b630027)

2026-05-10 13:29 PDT · bug-fix · 377.0s · $3.1778

Goal: Fix bunx @serranolabs.io/munchkins skills install failing with "could not determine executable to run for package" because the package declared no bin field.

Outcome: Added a bin entry mapping munchkins to ./src/index.ts in packages/munchkins/package.json, bumped the version from 0.1.1 to 0.1.2, and prepended a #!/usr/bin/env bun shebang to packages/munchkins/src/index.ts so the entrypoint is directly runnable. The existing if (import.meta.main) dispatch in src/index.ts remains the single source of truth — no wrapper file was introduced and no dispatch logic changed. Also extended the timeout on two integration-dispatch end-to-end tests in agent-builder.test.ts to 30s to accommodate the ~10 git subprocess invocations per run on slower/concurrent CI environments. Publishing was intentionally not performed; the user can run that as a separate manual step. Consider a follow-up to document bunx @serranolabs.io/munchkins … usage in the README.

Files changed:

packages/munchkins/package.json
packages/munchkins/src/index.ts
packages/munchkins-core/src/builder/agent-builder.test.ts

feat(agent-cli): wait for Claude rate-limit reset and retry once (c7e4fa8)

2026-05-10 13:00 PDT · feat-small · 742.8s · $4.0512

Goal: When the Claude CLI exits because the user's usage limit was hit, sleep until the reported reset time and retry the spawn exactly once instead of failing the whole pipeline.

Outcome: Switched the base AgentCLI.runJsonStream from stderr: "inherit" to stderr: "pipe" while forwarding bytes to process.stderr in real time, and added captured stderr to SpawnResult. Introduced file-local isLimitHit, parseResetTimestamp, and sleepUntil helpers in agent-cli.ts. ClaudeCLI.spawn() now inspects the result, sleeps until the parsed reset (unix-seconds or HH:MM, today/tomorrow), logs a single ⏳ Claude limit hit, waiting until <HH:MM:SS local> line to stderr, and retries the same spawn exactly once. CodexCLI.spawn is untouched.

How to test manually:

From the repo root, run the new unit tests to exercise all branches (unix-seconds retry, HH:MM, parse failure, abort mid-wait, double-limit no-third-spawn, stderr-only detection, out-of-range HH:MM):
```
bun test packages/munchkins-core/src/builder/agent-cli.test.ts
```
Expect every test in the ClaudeCLI rate-limit retry describe block to pass in well under a second.
Verify the repo-wide gates the deterministic loop runs:
```
bun run typecheck
bun run lint
bun run scenario
```
All three should be green.
Out-of-band manual check that the unit tests don't cover — confirm stderr is still forwarded live (the change from inherit to pipe is the riskiest part). In a throwaway script, drive a Claude spawn that prints to stderr and watch it appear in your terminal in real time:
```
bun -e "import { ClaudeCLI } from './packages/munchkins-core/src/builder/agent-cli.ts'; await new ClaudeCLI().spawn({ systemPrompt: '', userPrompt: 'say hi then exit', cwd: process.cwd() });"
```
You should see Claude's stderr stream appear as it's produced (not buffered until exit). If you have a real rate-limited account handy, trigger a limit and confirm you see the single ⏳ Claude limit hit, waiting until … line followed by exactly one retry after the reset.
Edge case — abort during wait. In a REPL, kick off a spawn with a fake limit message in stderr that points ~60s into the future via a stub of runJsonStream, then abort the controller after 50ms; spawn() should reject promptly with the abort reason and not fire a second spawn. (This is exactly what the aborted abortSignal during the wait test asserts; rerun that single test in watch mode if you want to poke at it: bun test packages/munchkins-core/src/builder/agent-cli.test.ts -t "aborted abortSignal".)

Files changed:

packages/munchkins-core/src/builder/agent-cli.ts
packages/munchkins-core/src/builder/agent-cli.test.ts

feat(munchkins-core): add resume subcommand for interrupted runs

2026-05-09 20:11 PDT · feat-small · 1532.9s · $15.7377

Goal: Add bun run munchkins resume [runId] so an interrupted run (rate limit, Ctrl-C, OOM) can pick up from the last completed step, including resuming the underlying Claude/Codex session.

Outcome: Each run now writes an incremental state.json to its .munchkins/runs/<slug>-<uuid>/ directory tracking phase, per-step status, and captured CLI session ids. SandboxFactory was reshaped into an object with create() + rehydrate(); gitWorktreeSandbox implements rehydrate() against an existing worktree with hard-fail preconditions for missing worktree/branch and a logged warning for dirty/advanced state. AgentBuilder exposes runFromState() (called by both fresh runs and the new resume orchestrator); ClaudeCLI/CodexCLI capture session_id from their JSONL streams and emit --resume <id> (Claude) or codex resume <id> exec (Codex) with a continue message when given a resumeSessionId. A new runResume(argv) orchestrator wired into packages/munchkins/src/index.ts handles --list, --latest, full runId, and unique-slug resolution; RunLog.resume(dir) replays events.jsonl so token/cost totals survive across the resume boundary.

How to test manually:

From the repo root, build/install once: bun install.
Confirm the new subcommand surfaces with no resumable runs: bun run munchkins resume --list — should print no resumable runs and exit 0. Same with no args: bun run munchkins resume.
Kick off a real bug-fix run against a throwaway change so it produces a state.json: bun run munchkins bug-fix --user-message="add a no-op comment to packages/munchkins-core/src/index.ts". Once you see the first agent step actually start (the banner prints [step 1/N agent]), interrupt with Ctrl-C.
Inspect the preserved run dir: ls .munchkins/runs/ then cat .munchkins/runs/<slug>-<uuid>/state.json — verify phase is still "steps" (not "done" / "failed"), step 0 has status: "in-progress", and (if Claude got far enough to emit its init event) sessionId is populated.
List resumables: bun run munchkins resume --list — should print a table row for that run with runId, agent, slug, started-at, phase, and completed/total step counts.
Resume by full id: bun run munchkins resume <slug>-<uuid>. Confirm in output that the worktree path is reused (no new agent/... branch is created), and that step 0 is re-attempted via claude --resume <id> (look for the continue message in verbose output by re-running with --verbose). When it finishes, state.json.phase should be "done".
Slug resolution edge case — happy path: with one resumable, run bun run munchkins resume <slug> (no uuid). Should resolve and run.
Slug resolution edge case — ambiguous: leave two interrupted runs with the same slug, run bun run munchkins resume <slug> — should exit 1 and print both full runIds in the error.
--latest: with multiple resumables, run bun run munchkins resume --latest and confirm it picks the most recent by startedAt.
Rehydrate hard-fail: from another resumable run dir, rm -rf the worktree it points at, then bun run munchkins resume <runId> — should exit 1 with Worktree at <path> no longer exists. Likewise delete its branch (git branch -D agent/...) for the deleted-branch failure.
Session-not-found fallback: edit a state.json's steps[0].sessionId to "definitely-bogus-id", then resume — Claude returns a session-not-found error and the builder should log [resume] session ... no longer available; restarting step with worktree-state hint. and re-spawn fresh with a git status / git diff --stat HEAD preamble baked into the system prompt.
Token accounting: after a resumed run completes, open .munchkins/runs/<slug>-<uuid>/summary.json and confirm tokens-in/out and cost reflect both the original run's events AND the resumed CLI calls (i.e. greater than what a fresh run alone would have produced).
Regression check on the unchanged path: run a fresh bun run munchkins bug-fix --user-message="..." to completion and confirm it still works end-to-end with no prompts about resume, no extra args, and produces the same shape of summary as before.
Automated tests cover the unit-level behavior — bun run test packages/munchkins-core/src/resume and bun run test packages/munchkins-core/src/sandbox/sandbox.test.ts — but step 11 above (real session-not-found against the live claude CLI) is the one out-of-band check the tests don't perform.

Files changed:

packages/munchkins-core/src/builder/agent-builder.ts
packages/munchkins-core/src/builder/agent-builder.test.ts
packages/munchkins-core/src/builder/agent-cli.ts
packages/munchkins-core/src/builder/agent-cli.test.ts
packages/munchkins-core/src/index.ts
packages/munchkins-core/src/resume/index.ts
packages/munchkins-core/src/resume/run-resume.ts
packages/munchkins-core/src/resume/run-resume.test.ts
packages/munchkins-core/src/resume/run-state.ts
packages/munchkins-core/src/resume/run-state.test.ts
packages/munchkins-core/src/resume/test-fixtures.ts
packages/munchkins-core/src/run-log.ts
packages/munchkins-core/src/run-log.test.ts
packages/munchkins-core/src/sandbox/index.ts
packages/munchkins-core/src/sandbox/sandbox.ts
packages/munchkins-core/src/sandbox/sandbox.test.ts
packages/munchkins/src/index.ts

fix(feat-small): swap new-surface section for manual-test recipe

2026-05-09 19:49 PDT · bug-fix · 261.2s · $1.6490

Goal: Replace the feat-small summary writer's New surface catalog with a How to test manually section so changelog entries give operators a concrete smoke-test recipe.

Outcome: Rewrote packages/munchkins/agents/feat-small/prompts/summary-writer.md. Deleted the What "new surface" means section and its category bullets, retargeted the framing paragraph at giving operators a manual smoke-test recipe, swapped the template's New surface block (and the now-redundant Lines added line) for a required How to test manually block with guidance on covering happy path plus an edge case and a fallback string for non-runtime features, and kept the Files changed block. Output contract, JSON envelope, no-headings-in-body rule, and feat(<scope>): <subject> commit format are unchanged. Default and refactor summary writers were not touched.

Files changed:

packages/munchkins/agents/feat-small/prompts/summary-writer.md

fix(builder): tolerate duplicate JSON envelope from summary writer

2026-05-09 19:17 PDT · bug-fix · 477.2s · $3.4697

Goal: Fix the summary writer JSON parser so it survives the model emitting the envelope twice in one response — a production failure mode (run agent-composition-integration-df872018) where the greedy regex spans both objects and JSON.parse chokes on the gap.

Outcome: Replaced the regex-based extraction in runSummaryWriter with a string-aware balanced-brace forward scan extracted into a new parseSummaryWriterJson helper. The scan enumerates every top-level {...} object in the response, then iterates them last-to-first and returns the first one that parses and has both commitMessage and markdown as strings. Trailing ``` fence handling and existing type checks are preserved; backward-compatible for the single-envelope case. Added 12 unit tests covering the duplicate-emit regression, prose preambles, fenced output, braces inside string literals, escaped quotes, missing/wrong-typed fields, and a realistic fixture.

Files changed:

packages/munchkins-core/src/builder/agent-builder.ts
packages/munchkins-core/src/builder/parse-summary-writer-json.ts
packages/munchkins-core/src/builder/parse-summary-writer-json.test.ts

Notes for future debuggers: The brace scanner is intentionally string-aware (tracks inString + isEscaped) so a JSON string value like "see {a, b, c}" or "escaped \"quote\"" doesn't desync the depth counter. If the scanner ever encounters an unbalanced {, it stops scanning rather than emitting a partial candidate. Failure reason distinguishes "no JSON object found" from "N candidate object(s) inspected" so harness logs point at the right diagnosis.

feat(core): pluggable integration strategies + agent composition

2026-05-09 16:00 PDT · feat-small · 1264.1s · $13.0129

Goal: Add pluggable integration strategies (integrateMerge / integratePR) and AgentBuilder composition (.thenRun()), wiring operator > author > default precedence and moving integration out of sandbox teardown into the run layer.

Outcome: Introduced IntegrationStrategy with integrateMerge (delegates to existing integrateBranch) and integratePR (rebase + push + open PR via gh/glab, with auto provider detection). Added .integrate(), .thenRun(), .setSandbox(), .rename(), .describe() builder methods; .thenRun() returns a new builder with steps concatenated and sandbox/summaryWriter/integration stripped per the S3 contract. Run-layer dispatch enforces precedence and surfaces a clear error for unknown modes; gitWorktreeSandbox.teardown is now cleanup-only. Added --integrate <mode> CLI flag, an example bugfix-then-refactor agent, and a composition scenario.

Note on this entry: the run's summary writer emitted the JSON envelope twice and the harness's greedy regex parser failed to parse it, so this entry is hand-authored from the summary writer's first emitted block (verbatim). The agent's actual work — the diff, the tests, and the deterministic checks — completed cleanly before the parser tripped. A follow-up bug-fix to the harness regex will harden the parser against duplicate-emit hiccups.

New surface:

Export: integrateMerge() (in packages/munchkins-core/src/integrate.ts)
Export: integratePR(opts?) (in packages/munchkins-core/src/integrate.ts)
Export: detectProvider(repoRoot, remote) (in packages/munchkins-core/src/integrate.ts)
Export: IntegrationStrategy, IntegrationContext, IntegrationResult, IntegratePROptions types (in packages/munchkins-core/src/integrate.ts)
Export: AgentBuilder.integrate(strategy) method (in packages/munchkins-core/src/builder/agent-builder.ts)
Export: AgentBuilder.thenRun(other) method (in packages/munchkins-core/src/builder/agent-builder.ts)
Export: AgentBuilder.setSandbox(factory) method (in packages/munchkins-core/src/builder/agent-builder.ts)
Export: AgentBuilder.rename(name) method (in packages/munchkins-core/src/builder/agent-builder.ts)
Export: AgentBuilder.describe(description) method (in packages/munchkins-core/src/builder/agent-builder.ts)
Export: RunLog.getAgentSummaryMarkdown(), RunLog.getAgentSummaryCommitMessage() (in packages/munchkins-core/src/run-log.ts)
Example agent: bugfix-then-refactor (in packages/munchkins/agents/bugfix-then-refactor/bugfix-then-refactor-agent.ts)
CLI flag: --integrate <mode> on every registered agent (registered in packages/munchkins-core/src/registry/registry.ts)
Env var: __MUNCHKINS_OPT_integrate
New scenario: scenarios/composition.ts
Other: PassOpts.prUrl field on RunLogger.pass() surfaces the PR URL in quiet and verbose output (in packages/munchkins-core/src/builder/run-logger.ts)
Removed: IntegrateContext export and TeardownContext.integrate field (sandbox teardown is now cleanup-only)

Files changed:

package.json
packages/munchkins-core/src/builder/agent-builder.test.ts (new)
packages/munchkins-core/src/builder/agent-builder.ts
packages/munchkins-core/src/builder/run-logger.test.ts (new)
packages/munchkins-core/src/builder/run-logger.ts
packages/munchkins-core/src/index.ts
packages/munchkins-core/src/integrate.test.ts
packages/munchkins-core/src/integrate.ts
packages/munchkins-core/src/registry/registry.ts
packages/munchkins-core/src/run-log.test.ts
packages/munchkins-core/src/run-log.ts
packages/munchkins-core/src/sandbox/index.ts
packages/munchkins-core/src/sandbox/sandbox.test.ts
packages/munchkins-core/src/sandbox/sandbox.ts
packages/munchkins/agents/bugfix-then-refactor/bugfix-then-refactor-agent.ts (new)
scenarios/composition.ts (new)

feat(scheduler): add cron support for AgentBuilder + daemon subcommand

2026-05-09 15:25 PDT · feat-small · 642.2s · $6.0585

Goal: Add per-agent cron scheduling via a new .cron() builder method and a bun run munchkins daemon entrypoint that arms timers per cronned builder and fires builder.run() at each tick.

Outcome: AgentBuilder now carries an optional CronConfig set via .cron(spec, { userMessage, verbosity }) and exposed via getCron(); calling .cron() twice throws naming the agent. A new scheduler/daemon.ts module collects cronned builders from a registry, renders a column-aligned startup table with ISO + humanized next-tick offsets via cron-parser, and arms one setTimeout per builder that resets per-tick env (__MUNCHKINS_OPT_verbose / __MUNCHKINS_OPT_thinking / __MUNCHKINS_OPT_userMessage) before each run, re-arming in finally. The munchkins entrypoint branches on process.argv[2] === "daemon" to invoke runDaemon() ahead of registry.cli(). No default agent is cronned; the API ships dormant.

New surface:

Export: AgentBuilder.prototype.cron(spec, opts) (in packages/munchkins-core/src/builder/agent-builder.ts)
Export: AgentBuilder.prototype.getCron() (in packages/munchkins-core/src/builder/agent-builder.ts)
Export: type Verbosity (in packages/munchkins-core/src/builder/agent-builder.ts, re-exported from builder/index.ts and src/index.ts)
Export: interface CronConfig (in packages/munchkins-core/src/builder/agent-builder.ts, re-exported from builder/index.ts and src/index.ts)
Export: runDaemon(opts?) (in packages/munchkins-core/src/scheduler/daemon.ts, re-exported from scheduler/index.ts and src/index.ts)
Export: applyTickEnv(cfg) (in packages/munchkins-core/src/scheduler/daemon.ts)
Export: collectCronnedBuilders(registry) (in packages/munchkins-core/src/scheduler/daemon.ts)
Export: interface CronnedBuilder (in packages/munchkins-core/src/scheduler/daemon.ts)
Export: interface RunDaemonOptions (in packages/munchkins-core/src/scheduler/daemon.ts)
CLI flag: daemon subcommand on the munchkins entrypoint (registered in packages/munchkins/src/index.ts as a pre-registry.cli() argv branch)
Env var: __MUNCHKINS_OPT_userMessage (set per tick by applyTickEnv)
Env var: __MUNCHKINS_OPT_verbose (set per tick when verbosity is "verbose")
Env var: __MUNCHKINS_OPT_thinking (set per tick when verbosity is "thinking")
New file: packages/munchkins-core/src/scheduler/daemon.ts
New file: packages/munchkins-core/src/scheduler/index.ts
New file: packages/munchkins-core/src/scheduler/daemon.test.ts
New package export path: @serranolabs.io/munchkins-core/scheduler (added in packages/munchkins-core/package.json)
Other: cron-parser@^5.5.0 added as a dependency of @serranolabs.io/munchkins-core

Lines added: +407 (across 8 files)

Files changed:

bun.lock
packages/munchkins-core/package.json
packages/munchkins-core/src/builder/agent-builder.ts
packages/munchkins-core/src/builder/index.ts
packages/munchkins-core/src/index.ts
packages/munchkins-core/src/scheduler/daemon.ts
packages/munchkins-core/src/scheduler/index.ts
packages/munchkins-core/src/scheduler/daemon.test.ts
packages/munchkins/src/index.ts

docs(summary-writer): forbid markdown headings in changelog body

2026-05-09 15:15 PDT · bug-fix · 321.1s · $2.0794

Goal: Stop the default summary-writer prompt from producing ## headings inside changelog entry bodies, which collide with the harness-emitted entry title and break the document hierarchy.

Outcome: Updated the Output contract section of packages/munchkins/agents/_shared/prompts/summary-writer.md to explicitly prohibit any Markdown headings (#, ##, ###, etc.) inside the markdown field. Promoted the bold inline labels (**Goal:**, **Outcome:**, **Files changed:**) from a suggested skeleton to a required body shape, and added a side-by-side correct/wrong example so the contrast is visual. The JSON output contract (commitMessage + markdown keys, no code fences) is unchanged, and the harness-side assembly in RunLog.prependChangelogIn was not touched. Per-agent summary-writer prompts under feat-small and refactor were left alone per the task constraints (they were out of scope unless they shared the defect).

Files changed:

packages/munchkins/agents/_shared/prompts/summary-writer.md

Future changelog entries produced by agents using this default prompt should contain zero #-prefixed lines in the body while still carrying the bold inline labels for Goal / Outcome / Files changed.

fix(munchkins-core): detect merge-fixer progress via working-tree content

2026-05-09 15:04 PDT · bug-fix · 451.4s · $2.6168

Goal

Fix the merge-fixer harness in integrate.ts so it stops misclassifying every real conflict resolution as "no progress."

Outcome

Replaced the index-based post-fixer progress check with a content-based check using git diff --check. The new logic detects leftover conflict markers in the working tree, bails out if the fixer wrote markers to files outside the conflict set, fails with a no-progress reason only when every conflicted file still has markers, and stages only files verified clean (per-file git add, never git add -A). Added a small abortAndFail helper to deduplicate the bail-out paths in rebaseAndResolve. Introduced packages/munchkins-core/src/integrate.test.ts with five real-integrateBranch tests covering the happy path, no-progress, partial-progress (regression test that proves the outer loop re-prompts the fixer), CLI failure, and the clean-rebase no-spawn case.

Files changed

packages/munchkins-core/src/integrate.ts — swap index-based stillConflicted check for filesWithLeftoverMarkers (new helper using git diff --check); add stray-file guard; switch staging to per-file git add; extract repeated abortRebase + return into a local abortAndFail helper.
packages/munchkins-core/src/integrate.test.ts — new file. mkdtemp + git init -b main repos, gitEnv() helper with TEST_GIT_IDENTITY, StubFixerCLI with constructor-injected handler and FailIfSpawnedCLI, plus shared setupSingleFileConflict / setupTwoFileConflict setup helpers driving the five required scenarios.

Notes for future readers

The partial-progress test (#3) is the load-bearing regression: it asserts cli.invocations === 2 and fixerIters === 2, which is only achievable if the harness correctly recognizes per-file forward progress and re-enters the fixer for the remaining unresolved file.
listConflictedFiles is intentionally retained for the initial per-iteration enumeration before the fixer runs; only the post-fixer check moved to a content-based detector.

refactor(agent-builder): extract RunLogger to centralize verbose/quiet formatting

2026-05-09 14:25 PDT · refactor · 590.3s · $3.3619

Goal: Extract every verbose/quiet branch site in AgentBuilder.run() (and its runAgent / runDeterministic / runSummaryWriter helpers) into a single RunLogger class so run() orchestrates and RunLogger formats — byte-identical output in both modes.

Outcome: Created packages/munchkins-core/src/builder/run-logger.ts housing the new RunLogger class plus the C color table, banner(), and printInvocation() helpers (moved out of agent-builder.ts so there's a single home for terminal formatting). AgentBuilder imports them back, drops the inline if (verbose) … else … blocks at all twelve call sites, and threads a single RunLogger instance into the helpers in place of the bare verbose: boolean parameter. streamOutput stays where it was — it's a separate concern (Claude streaming) — and the env reads at the top of run() are untouched.

Refactor type: extraction

Lines changed:

File	Before	After	Δ
packages/munchkins-core/src/builder/agent-builder.ts	642	525	−117
packages/munchkins-core/src/builder/run-logger.ts	0	200	+200
packages/munchkins-core/src/index.ts	46	46	0

Total: 688 → 771 (Δ +83)

Files changed:

packages/munchkins-core/src/builder/agent-builder.ts
packages/munchkins-core/src/builder/run-logger.ts
packages/munchkins-core/src/index.ts

Call sites that now share the extracted helpers:

RunLogger.stepResultOk() — called by both runAgent and runSummaryWriter, replacing the duplicated quiet-mode " ok (Xs, in→out)\n" line that previously existed inline in each.
RunLogger.starting() / stepBanner() / pass() / fail() / logDir() / integrationLine() — collapse the six verbose-vs-quiet branch pairs in AgentBuilder.run() into single calls.
RunLogger.summaryWriterEmptyDiff() / summaryWriterStart() — collapse the two branch pairs in runSummaryWriter.
RunLogger.agentStepStart() — collapses the single branch pair in runAgent.
RunLogger.deterministicIterationHeader() / deterministicCommand() / deterministicCommandOutput() / deterministicQuietSummary() / fixerStart() / fixerResult() — collapse the six branch pairs in runDeterministic.
banner() and printInvocation() — now defined once in run-logger.ts and re-imported by agent-builder.ts instead of being module-local helpers.

refactor(munchkins/agents): extract getAgentPromptsDir helper

2026-05-09 14:12 PDT · refactor · 243.6s · $1.6386

Goal: Eliminate the 4× duplicated dirname(fileURLToPath(import.meta.url)) + join("prompts") incantation across bugfix-agent.ts, refactor-agent.ts, feat-small-agent.ts, and presets.ts by adding a single getAgentPromptsDir(importUrl) helper to presets.ts.

Outcome: Added getAgentPromptsDir(importUrl: string) to packages/munchkins/agents/_shared/presets.ts and re-exported it. The shared SHARED_PROMPTS constant and all three agent files (bugfix, feat-small, refactor) now call the helper instead of recomputing dirname(fileURLToPath(import.meta.url)) inline. The node:url import was dropped from the three agent files; only join is still pulled from node:path where needed alongside the helper. The trailing import-order delta in packages/munchkins-core/src/index.ts is incidental cleanup from the biome lint pass that ran with this refactor.

Refactor type: extraction

Lines changed:

File	Before	After	Δ
packages/munchkins-core/src/index.ts	43	43	0
packages/munchkins/agents/_shared/presets.ts	26	30	+4
packages/munchkins/agents/bugfix/bugfix-agent.ts	39	39	0
packages/munchkins/agents/feat-small/feat-small-agent.ts	46	46	0
packages/munchkins/agents/refactor/refactor-agent.ts	28	32	+4

Total: 182 → 190 (Δ +8)

Files changed:

packages/munchkins-core/src/index.ts
packages/munchkins/agents/_shared/presets.ts
packages/munchkins/agents/bugfix/bugfix-agent.ts
packages/munchkins/agents/feat-small/feat-small-agent.ts
packages/munchkins/agents/refactor/refactor-agent.ts

Call sites now sharing getAgentPromptsDir:

packages/munchkins/agents/_shared/presets.ts — defines and consumes it for SHARED_PROMPTS
packages/munchkins/agents/bugfix/bugfix-agent.ts — const PROMPTS = getAgentPromptsDir(import.meta.url)
packages/munchkins/agents/feat-small/feat-small-agent.ts — const PROMPTS = getAgentPromptsDir(import.meta.url)
packages/munchkins/agents/refactor/refactor-agent.ts — const PROMPTS = getAgentPromptsDir(import.meta.url)

Net line count grew by 8 because the helper is defined once and each agent still needs an import line for it; the win is that the dirname(fileURLToPath(...)) pattern is no longer repeated and there is now a single place to change how agent prompt directories are resolved.

docs: add README with concise repo intro and usage

2026-05-08 21:21 PDT · feat-small · 300.9s · $1.7630

Goal: Add a concise, compelling README with Usage upfront, then push.

Outcome: Created a new top-level README.md that opens with a one-paragraph pitch, then leads with Usage (install + invocation + examples), followed by default agents, backend selection, per-run lifecycle, repo layout, a build-your-own-agent snippet, requirements, and a pointer to AGENTS.md. Also tightened docs/tsconfig.json to use @types/bun instead of node.

New surface:

New file: README.md

Lines added: +87 (across 2 files)

Files changed:

README.md
docs/tsconfig.json

feat(core): add switchable agent CLI backend (claude/codex)

2026-05-08 20:59 PDT · feat-small · 671.0s · $6.1088

Goal: Add a process-wide backend switch so operators can run any agent against either the claude CLI (default) or the codex CLI, selected by --cli flag or MUNCHKINS_CLI env var.

Outcome: Introduced an abstract AgentCLI base class with ClaudeCLI and CodexCLI subclasses behind a single AgentCLI.fromEnv() seam. The existing spawnClaude export collapses to a one-liner that delegates to AgentCLI.fromEnv().spawn(), preserving the harness mock seam and all 3 internal call sites unchanged. Cost tracking becomes optional end-to-end (AgentUsage.costUsd?, RunSummary.costUsd?, RunLog.getCostUsd() returns number | undefined) so Codex's missing cost data renders as — honestly instead of being faked as $0.0000. Codex's missing system-prompt flag is handled by prepending ## System\n…\n\n## Task\n… as a single positional argument.

New surface:

Export: AgentCLI abstract class (in packages/munchkins-core/src/builder/agent-cli.ts)
Export: ClaudeCLI class with buildArgs() and spawn() (in packages/munchkins-core/src/builder/agent-cli.ts)
Export: CodexCLI class with buildArgs() and spawn() (in packages/munchkins-core/src/builder/agent-cli.ts)
Export: AgentCLI.fromEnv() static factory (in packages/munchkins-core/src/builder/agent-cli.ts)
Export: types SpawnOptions, SpawnResult, AgentUsage, AgentCLIName (in packages/munchkins-core/src/builder/agent-cli.ts, re-exported from packages/munchkins-core/src/builder/index.ts)
CLI flag: --cli <cli> on every registered agent subcommand (registered in packages/munchkins-core/src/registry/registry.ts)
Env var: MUNCHKINS_CLI (public; read by AgentCLI.fromEnv())
Env var: __MUNCHKINS_OPT_cli (internal flag-bridge set by the registry; takes priority over MUNCHKINS_CLI)
New file: packages/munchkins-core/src/builder/agent-cli.ts
New file: packages/munchkins-core/src/builder/agent-cli.test.ts

Lines added: +484 (across 8 files)

Files changed:

AGENTS.md
packages/munchkins-core/src/builder/agent-builder.ts
packages/munchkins-core/src/builder/agent-cli.ts (new)
packages/munchkins-core/src/builder/agent-cli.test.ts (new)
packages/munchkins-core/src/builder/index.ts
packages/munchkins-core/src/builder/spawn-claude.ts
packages/munchkins-core/src/registry/registry.ts
packages/munchkins-core/src/registry/registry.test.ts
packages/munchkins-core/src/run-log.ts
packages/munchkins-core/src/run-log.test.ts

feat(core): human-readable run names from LLM slug

2026-05-08 20:29 PDT · feat-small · 1077.4s · $9.3383

Goal: Replace timestamp-based run-log directory names and worktree branch names with LLM-generated, human-readable slugs derived from the user's task description.

Outcome: Added a slug-derivation pipeline (Haiku LLM call with retry + deterministic kebab fallback) that feeds both RunLog directory names and worktree branch names. Run dirs are now <slug>-<uuid8> and branches are agent/<slug>-<uuid8>, with the slug capped at 30 chars. Slug derivation runs in parallel with sandbox creation and the worktree branch is renamed once the slug resolves; fallbacks are recorded as slug-fallback events in events.jsonl. The scenario harness recognizes the new naming via a Haiku-aware mock branch.

New surface:

Export: sanitize(raw) (in packages/munchkins-core/src/builder/slug.ts)
Export: deriveSlugDeterministic(text) (in packages/munchkins-core/src/builder/slug.ts)
Export: getSlugWithRetry(userMessage, opts?) (in packages/munchkins-core/src/builder/slug.ts)
Export: SLUG_MAX constant (in packages/munchkins-core/src/builder/slug.ts)
Export: SlugResult type (in packages/munchkins-core/src/builder/slug.ts)
Export: SlugFallback type (in packages/munchkins-core/src/builder/slug.ts)
Export: renameBranch(oldBranch, newBranch, repoRoot) (in packages/munchkins-core/src/worktree.ts)
Export: RunLog.recordEvent(event) public method (in packages/munchkins-core/src/run-log.ts)
Export: getSlugOutput() (in scenarios/lib/mock-spawn-claude.ts)
Export: New slug?: string option on RunLog constructor (in packages/munchkins-core/src/run-log.ts)
Export: New optional branchName parameter on createWorktree (in packages/munchkins-core/src/worktree.ts)
Export: New model, disallowedTools, abortSignal fields on SpawnClaudeOptions (in packages/munchkins-core/src/builder/spawn-claude.ts)
New file: packages/munchkins-core/src/builder/slug.ts
New file: packages/munchkins-core/src/builder/slug.test.ts
New file: packages/munchkins-core/src/run-log.test.ts
New file: packages/munchkins-core/src/worktree.test.ts
Other: New slug-fallback event type written to events.jsonl (recorded from agent-builder.ts)
Other: claude CLI now invoked with --model and --disallowedTools flags when those options are set (in packages/munchkins-core/src/builder/spawn-claude.ts)

Lines added: +362 (across 12 files)

Files changed:

packages/munchkins-core/src/builder/agent-builder.ts
packages/munchkins-core/src/builder/index.ts
packages/munchkins-core/src/builder/slug.ts
packages/munchkins-core/src/builder/slug.test.ts
packages/munchkins-core/src/builder/spawn-claude.ts
packages/munchkins-core/src/index.ts
packages/munchkins-core/src/run-log.ts
packages/munchkins-core/src/run-log.test.ts
packages/munchkins-core/src/sandbox/sandbox.ts
packages/munchkins-core/src/sandbox/sandbox.test.ts
packages/munchkins-core/src/worktree.ts
packages/munchkins-core/src/worktree.test.ts
scenarios/index.ts
scenarios/lib/mock-spawn-claude.ts

refactor: extract duplicated helpers across builder, run-log, scenarios

2026-05-08 20:11 PDT · refactor · 408.3s · $3.3636

Goal: Fix any refactoring opportunities found across the codebase.

Outcome: Pulled four sets of duplicated inline logic into named helpers/constants. In agent-builder.ts, the user-message resolution and summary-writer prompt construction were duplicated across _buildSummaryWriterUserPrompt-equivalent call sites; these are now resolveOriginalGoal() and buildSummaryWriterUserPrompt(). In run-log.ts, the command-entry log formatter and env-path resolution logic were each duplicated; extracted as formatCommandEntries() and resolveEnvPath(). In sandbox.test.ts, the git identity env block was inlined twice; extracted to TEST_GIT_IDENTITY constant and gitEnv() helper. In scenarios/index.ts, six near-identical fail-result objects were collapsed into a single failResult() factory closure. Net line count drops, but the primary value is single source of truth for each pattern.

Refactor type: extraction

Lines changed:

File	Before	After	Δ
packages/munchkins-core/src/builder/agent-builder.ts	656	637	−19
packages/munchkins-core/src/run-log.ts	331	332	+1
packages/munchkins-core/src/sandbox/sandbox.test.ts	182	181	−1
scenarios/index.ts	358	311	−47

Total: 1527 → 1461 (Δ −66)

Files changed:

packages/munchkins-core/src/builder/agent-builder.ts
packages/munchkins-core/src/run-log.ts
packages/munchkins-core/src/sandbox/sandbox.test.ts
scenarios/index.ts

Extracted helpers and their call sites:

resolveOriginalGoal(repoRoot) in agent-builder.ts — called from the runtime summary-writer phase (~line 317) and the resolved-prompt preview phase (~line 442).
buildSummaryWriterUserPrompt(originalGoal, diffSection) in agent-builder.ts — called from the same two phases as above.
formatCommandEntries(entries) in run-log.ts — called from recordDeterministic (~line 206) and recordFinalize (~line 244).
resolveEnvPath(envValue, fallback, repoRoot) in run-log.ts — called from the RunLog constructor (run-log dir) and _prependChangelog (changelog path).
TEST_GIT_IDENTITY constant + gitEnv() helper in sandbox.test.ts — used by createRepo() and commit().
failResult(phase, message, opts?) closure in scenarios/index.ts — replaces six inline fail-shaped returns covering registry-miss, agent failure, mock-call audit, claude-spawn audit, cleanup assertion, artifact assertion, and the catch-all setup error.

feat(cli): add --thinking flag for mid verbosity level

2026-05-08 19:58 PDT · feat-small · 582.6s · $4.0524

Goal: Add a middle verbosity level called --thinking that sits between the default (terse) and --verbose (full), streaming Claude output without the boxed prompt prefaces.

Outcome: Registered a new --thinking Commander option on every agent subcommand that sets __MUNCHKINS_OPT_thinking=true. AgentBuilder.run() now reads both verbose and thinking env vars and computes a streamOutput flag (verbose || thinking) that is threaded through runAgent, runDeterministic, and the writer phase to control whether spawnClaude streams. Boxed printInvocation output, phase banners, and per-command stdout remain gated on the existing verbose flag, so --thinking only unlocks streaming. Updated --verbose help text to mention the three levels and added registry tests covering option registration and env-var wiring.

New surface:

CLI flag: --thinking on every registered agent subcommand (registered in packages/munchkins-core/src/registry/registry.ts)
Env var: __MUNCHKINS_OPT_thinking
New file: packages/munchkins-core/src/registry/registry.test.ts

Lines added: +124 (across 4 files)

Files changed:

packages/munchkins-core/src/builder/agent-builder.ts
packages/munchkins-core/src/registry/registry.ts
packages/munchkins-core/src/registry/registry.test.ts
packages/munchkins-core/src/sandbox/sandbox.ts

refactor(run-log): extract _writeClaudeCall helper

2026-05-08 19:02 PDT · refactor · 267.4s · $1.6061

Goal: Deduplicate three near-identical Claude-call writers in RunLog (agentStep, summaryStep, fixerInvocation) by extracting a private helper.

Outcome: Added _writeClaudeCall(prefix, kind, systemPrompt, userPrompt, response, exitCode, durationMs, extra) that writes the *.system.md / *.user.md / *.response.txt artifacts, increments claudeCallCount, and emits the corresponding event with systemBytes / userBytes / responseBytes. Each public method now computes only its own prefix and per-kind extra fields (stepIndex, iteration) and delegates to the helper, while preserving its own counter bumps (agentStepCount, fixerInvocationCount). Behavior, file names, and event payload shapes are unchanged.

Files changed:

packages/munchkins-core/src/run-log.ts — introduced _writeClaudeCall helper; rewrote agentStep, summaryStep, and fixerInvocation to delegate to it; consolidated the claudeCallCount increment into the helper.

No agent runs have been recorded yet. The first successful run of bun run munchkins <agent> will prepend an entry above.

#Changelog

#feat(munchkins): launch agent runs inside cmux workspace when available (86ddc5a)

#feat(munchkins-core): add Prompt.withSkill() helper (07441da)

#feat(munchkins): add SKILL.md discovery surface for default agents (bd12fe3)

#feat(docs): re-orient onboarding around the new-munchkin skill (deb5605)

#docs(pages): add user-facing agent guide organized by agent (a01c8fb)

#fix(munchkins): make package bunx-executable via bin entry (b630027)

#feat(agent-cli): wait for Claude rate-limit reset and retry once (c7e4fa8)

#feat(munchkins-core): add resume subcommand for interrupted runs

#fix(feat-small): swap new-surface section for manual-test recipe

#fix(builder): tolerate duplicate JSON envelope from summary writer

#feat(core): pluggable integration strategies + agent composition

#feat(scheduler): add cron support for AgentBuilder + daemon subcommand

#docs(summary-writer): forbid markdown headings in changelog body

#fix(munchkins-core): detect merge-fixer progress via working-tree content

#Goal

#Outcome

#Files changed

#Notes for future readers

#refactor(agent-builder): extract RunLogger to centralize verbose/quiet formatting

#refactor(munchkins/agents): extract getAgentPromptsDir helper

#docs: add README with concise repo intro and usage

#feat(core): add switchable agent CLI backend (claude/codex)

#feat(core): human-readable run names from LLM slug

#refactor: extract duplicated helpers across builder, run-log, scenarios

#feat(cli): add --thinking flag for mid verbosity level

#refactor(run-log): extract _writeClaudeCall helper

Changelog

feat(munchkins): launch agent runs inside cmux workspace when available (86ddc5a)

feat(munchkins-core): add Prompt.withSkill() helper (07441da)

feat(munchkins): add SKILL.md discovery surface for default agents (bd12fe3)

feat(docs): re-orient onboarding around the new-munchkin skill (deb5605)

docs(pages): add user-facing agent guide organized by agent (a01c8fb)

fix(munchkins): make package bunx-executable via bin entry (b630027)

feat(agent-cli): wait for Claude rate-limit reset and retry once (c7e4fa8)

feat(munchkins-core): add resume subcommand for interrupted runs

fix(feat-small): swap new-surface section for manual-test recipe

fix(builder): tolerate duplicate JSON envelope from summary writer

feat(core): pluggable integration strategies + agent composition

feat(scheduler): add cron support for AgentBuilder + daemon subcommand

docs(summary-writer): forbid markdown headings in changelog body

fix(munchkins-core): detect merge-fixer progress via working-tree content

Goal

Outcome

Files changed

Notes for future readers

refactor(agent-builder): extract RunLogger to centralize verbose/quiet formatting

refactor(munchkins/agents): extract getAgentPromptsDir helper

docs: add README with concise repo intro and usage

feat(core): add switchable agent CLI backend (claude/codex)

feat(core): human-readable run names from LLM slug

refactor: extract duplicated helpers across builder, run-log, scenarios

feat(cli): add --thinking flag for mid verbosity level

refactor(run-log): extract _writeClaudeCall helper