v0.3.1 — zero-install, self-documenting CLI

odw is now zero install. Nothing is scaffolded into your project — the binary is self-documenting, so any AI (Claude, codex, or otherwise) reads the CLI and can use it.

Highlights

• odw guide — a single self-contained authoring + run guide (what odw is, the full API, how to run, the gotchas). odw --help leads with it; odw spec adds the TypeScript types, odw contract the authoring contract.
• Removed the install model — odw init and the pack-management commands (validate/agents/workflows/graphs/evidence) and the installed pack (skill, agents, slash commands, starter workflows, schema files) are gone. ~1000 lines lighter. doctor now reports node + pandacode + runtimes only.
• The embedded runtime, HTML report assets, and authoring contract stay inside the binary.

Validation

A fresh agent handed only the CLI (no source, no README) went from zero knowledge to a working multi-node workflow run (parallel + schema + pipeline + HTML report) using odw guide alone — proving the self-documenting design. It surfaced one real gap (now fixed): the guide now documents that a --backend mock dry run returns a status object (not final text) and that schema nodes always "fail" under mock — use mock to verify graph shape, real runs for node outputs.

Gate green: cargo build/test/clippy + selftest 52/52. odw exec / report / runs / doctor / spec / contract / capabilities unchanged.

v0.3.0 — dogfood hardening round 2

Continued autonomous dogfooding (build real projects, run varied real workflows, probe edge cases) since v0.2.0 — 19 commits of fixes, all gate-green (cargo build/test/clippy + 52/52 parity selftest). Binaries now correctly report their version.

Security & robustness

• Report XSS: a node prompt containing </script> could execute when the offline HTML report opened — embedded JSON is now <-escaped.
• Unloadable schema now fails fast with a clear non-retryable schema_load_error instead of burning retries as a confusing mismatch.
• Non-serializable workflow return (circular/BigInt/function) → a clean structured result_not_serializable failure instead of an opaque crash.
• pandacode: session name sanitized (path-traversal), atomic session writes, no tmux leak on a failed resume turn.

Executor / PandaCode

• bamboo: infer the provider from the model when --provider is omitted — every non-deepseek model previously misrouted to deepseek (400).
• Unchanged worktree nodes keep a consistent {text, worktree} shape, so result.worktree.changed is always inspectable.

Examples & docs (honest parity)

• Examples 03/05/06 fixed to actually work on real runs / read cleanly in mock dry-runs.
• Documented the real divergences from the built-in Workflow: budget counts total (not output) tokens; schema retry is opt-in (maxAttempts:1); a new "where the built-in is better" section; --timeout granularity.
• Removed stale references to the converged-away odw mcp / odw run surfaces.

The runtime contract matches the built-in Workflow tool on the behaviors covered by scripts/selftest.mjs (concurrency caps, determinism guards, parallel/pipeline semantics, budget, worktree, workflow() nesting).

v0.2.0 — dogfood-hardened

27 atomic, gate-green, CI-green commits from an autonomous dogfooding loop (real projects, 100-node scale runs, head-to-head vs the built-in tool).

Fixes

• claude runtime: strip the PANDACODE_DONE completion marker from results; reliably capture the final reply via the hook event log; report token usage from the session transcript (budget now counts claude).
• observability: backfill the executor-resolved model (no more inherit); report shows a per-model token breakdown that reconciles to the total.
• budget accrues on every attempt (retries/failures counted), documented as best-effort under concurrency.
• codex node timeout floors at 10 min so real builds aren't truncated; codex defaults to full access (workspace-write blocks the network and breaks npm/pip/cargo install).
• security: block a symlink escape in the agent write tools (create_dirs path); harden runs show run-id resolution against path traversal.
• out-of-box: auto-detect a sibling pandacode binary; doctor checks Node.js; concise pack status; relative --script falls back to cwd.

Docs/examples: heterogeneous-models + build-a-project examples; honest 'when the built-in tool is better'; corrected bamboo guidance (it's a coding agent — route prose/structured nodes to claude/codex).

Validated: 100 real parallel nodes (concurrency cap held); Python/Rust/Node projects built end-to-end; nested workflows + worktree concurrency + resume; ~3x cheaper than the built-in tool on the same fan-out.

v0.1.0 — initial public release

First public release. odw (orchestration runtime) + pandacode (codex/claude/bamboo executor) shipped as one Cargo workspace. Clone → ./install.sh → odw doctor → run. Auto HTML execution-graph reports, deterministic resume, worktree isolation, token budgets.