Skip to content
Chimera

claw-code

Origin: ultraworkers/claw-code (2026, MIT) — a public rewrite that exists downstream of a 2026 leak of a commercial agent harness’s source; its README disclaims affiliation with Anthropic and “does not claim ownership of the original Claude Code source material” Loop type: streaming tool-call conversation loop (a ConversationRuntime in the canonical Rust workspace); loop internals are not documented in the upstream READMEs — replicated in Chimera as ReAct with a 25-step ceiling plus opt-in bounded rerun Primary surface: terminal REPL, one-shot prompt, and direct CLI subcommands (rusty-claude-cli crate: “REPL, one-shot prompt, direct CLI subcommands, streaming display”) Chimera primitive: chimera/badger/ (verified at 2507d0c)

The repo describes itself as “the public Rust implementation of the claw CLI agent harness” and frames itself as an “agent-managed museum exhibit.” The thesis Chimera’s parity notes carry forward from it: better harness tools, not merely storing the archive — a clean rewrite is worth more than a museum-quality copy of the leaked snapshot. A naming note: Chimera’s badger trees (chimera/badger/, docs/badger/, tests/badger/, the badger site docs) never use this page’s title, by policy; they say “the upstream” or “the harness-rewrite tradition.” This Field Guide page, like the inspirations index, is the designated factual reference — do not copy its title into badger-scoped files.

Loop

Section titled “Loop”

The upstream READMEs document the harness around the loop — crates, tools, slash palette, parity audit — but not the turn mechanics themselves. The diagram below is the posture as made executable in chimera/badger/ (cli.py + rerun.py), which is the loop this page can verify firsthand:

flowchart TD
    A[User prompt] --> B[ReAct step: reason, call tool, observe]
    B --> C{Model stops calling tools, or 25-step ceiling?}
    C -- not yet --> B
    C -- turn over --> D[Scan trajectory for failure markers]
    D -- no markers --> E[Return result as-is]
    D -- markers, rerun budget left --> F[Reset conversation; prepend refined directive naming the markers]
    F --> B
    D -- markers, budget exhausted --> G[Return last failed result]

The distinguishing move is that recovery happens across attempts, not within one. Where sibling agents extend a long trajectory to dig out of a failure, this posture caps the trajectory (25 steps vs 50 for Chimera’s other CLIs) and spends the saved budget on a fresh attempt with a refined prompt — total attempts = 1 + --max-reruns (default 2).

Tool Set

Section titled “Tool Set”

The tools crate’s own enumeration (“tool specs + execution”), mapped to badger’s surface — badger ships Chimera’s shared tool set, restrictable with --allowed-tools:

Upstream toolPurposebadger equivalent
Bashshell executionbash
ReadFile / WriteFile / EditFileread, create/overwrite, in-place editread / write / edit
GlobSearch / GrepSearchfilename and content searchlist_files / search
WebSearch / WebFetchweb search and page fetchweb_fetch (fetch only)
Agentsub-agent spawndelegate
TodoWritetodo list maintenancetodo
NotebookEditnotebook editing—
Skillskill invocationskills discovery in the shared REPL
ToolSearchtool discovery—

Per-tool constraints are not documented in the workspace README; gating lives in the runtime crate, which owns the “permission policy.” Badger’s analog is the cross-CLI --permission-mode flag (read-only / suggest / auto / yolo / strict, default suggest).

Prompt Strategy

Section titled “Prompt Strategy”
  • The upstream READMEs do not publish the system prompt. What they document is the surface around it: model aliases (opus / sonnet / haiku, default claude-opus-4-7), a slash palette (/help, /status, /sandbox, /cost, /resume, /mcp, /agents, /skills, /doctor, /plugin, /subagent), and tab completion that “expands slash commands, model aliases, permission modes, and recent session IDs.”
  • The replica’s system prompt is a three-sentence posture statement: “You are Badger, a Chimera coding agent in the harness-rewrite tradition. Plan briefly, act with a tight tool budget, and verify before declaring success.” No few-shot examples.
  • On rerun, the next attempt’s prompt is the original prefixed by a directive that names the detected markers and ends “Do not claim done until verification passes.”
  • Edit format: tool-call file edits (EditFile / WriteFile upstream; read–edit–write in the replica). No diff dialect in prompts.

Context Strategy

Section titled “Context Strategy”
  • Upstream: the runtime crate owns “config loading, session persistence, permission policy, MCP client lifecycle”; /resume plus recent-session tab completion imply persistent, resumable sessions. Compaction behavior is not documented in the READMEs.
  • Replica: every run persists to an event-sourced log at ~/.chimera/eventlog/badger-<utc>-<uuid8>/, inspectable via sessions list/show/cost and exportable via share. Compaction is manual (/compact), history reset is manual (/clear).
  • The 25-step ceiling bounds context growth by construction; a rerun resets the conversation rather than appending to it, so recovery comes from a refined fresh context, not a longer one.

Termination Heuristic

Section titled “Termination Heuristic”

The upstream READMEs do not document termination. The replica makes the posture’s termination rules explicit:

  • A turn ends when the model stops calling tools or hits the 25-step ceiling.
  • With --rerun-on-failure set, the finished trajectory is scanned against a deliberately conservative marker list: pytest FAILED lines and failure summaries, SyntaxError, Python tracebacks, JavaScript SyntaxError … Unexpected, Rust error[E…] compile errors, node AssertionError, and explicit BUILD FAILED / TESTS FAILED lines.
  • Markers + budget remaining → reset and rerun with the refined prompt; markers with budget exhausted → return the last failed result.
  • No markers → return as-is, even if the agent reported failure — rerunning without an actionable marker would be guesswork. False negatives are preferred over false positives: rerun cost is treated as a real budget, not a free safety net.

Notable Quirks

Section titled “Notable Quirks”
  • Museum-exhibit framing. “This is not meant to be hand-operated like a normal product repo. It is an agent-managed exhibit: the harnesses plan, execute, verify, label, and preserve the artifact while the crabs keep the tank running.” The README adds, plainly: “Claw Code is not the serious production project here.”
  • The wrong-crate trap. cargo install claw-code installs a deprecated crates.io stub that only prints “claw-code has been renamed to agent-code.” The repo is build-from-source; the renamed upstream binary installs as agent-code.
  • Parity as a first-class artifact. PARITY.md is “the current Rust-port checkpoint”; a mock parity harness ships 10 coverage scenarios (mock_parity_scenarios.json, scripts/run_mock_parity_diff.py) and an entire mock-anthropic-service crate exists for “deterministic /v1/messages mock for CLI parity tests.”
  • Python demoted to reference. The src/ + tests/ Python workspace is the “companion Python/reference workspace and audit helpers; not the primary runtime surface” — the Rust workspace (9 crates) is canonical.
  • Affiliation disclaimers up front. “This repository is not affiliated with, endorsed by, or maintained by Anthropic,” and it “does not claim ownership of the original Claude Code source material.”

In Chimera

Section titled “In Chimera”

Badger (chimera badger, alias chimera strict) reimplements the harness-rewrite posture on Chimera’s shared primitives. Adopted:

  • Strict defaults — --max-steps defaults to 25 vs 50 for the six sibling CLIs; --allowed-tools narrows the surface; named --profile presets (strict / balanced / yolo) bundle permission mode, step budget, and rerun stance into one flag.
  • The PARITY pattern, made live — chimera badger parity --against PARITY.md loads a declared schema (JSON, YAML-ish, or fenced-Markdown block) and diffs it against the running agent’s tools, max-steps, slash commands, default model, and rerun flag. The diff is asymmetric (live extras are informational; only declared fields fail) and exits 0 on match, 1 on drift, 2 on usage error — CI-enforceable, like the upstream’s parity audit.
  • Rerun-on-failure as a first-class flag — --rerun-on-failure / --max-reruns plus the /rerun REPL command, formalizing the tradition’s verify-before-claiming-done discipline.
  • Anthropic-first provider chain — --model > $BADGER_MODEL > Anthropic (claude-sonnet-4-6) > OpenAI > OpenRouter > Ollama, with a friendly error listing every supported env var.

Diverged: Python on the shared Agent / ReAct / eventlog substrate rather than a Rust workspace; parity checks a live agent against an operator schema rather than tracking port progress against a reference codebase; the default model is claude-sonnet-4-6 rather than claude-opus-4-7; serve / agents / bench are stubs that defer to sibling CLIs; and rerun-on-failure has no documented upstream flag equivalent — it is badger’s executable reading of the posture. Full surface-by-surface status: parity matrix.

References

Section titled “References”