Command reference
このページはまだ翻訳されていないため、英語の原文を表示しています。
Every command the agent has, read from the CLI itself.
Generated from the product's own command surface and regenerated by its CI, so a renamed flag cannot survive here. The help text is the text you get from --help — it is the code's own words and is deliberately not translated.
Documenting version 0.51.0 · 142 commands
Set up and check
chimera initFirst-run setup: create .env, set a provider key, and point you at a real example.chimera doctorCheck the environment and configuration. With --fix, repair safe setup issues. `--probe` is off by default and that is deliberate: `doctor` should stay instant, offline and free. What it buys when you ask for it is the difference between a claim and a measurement — "Ready" below is an assertion about the NAME of an environment variable, so a revoked key, an account with no credit, or a value pasted with a trailing space all pass it and fail on the first real call. The argument for measuring is already written in `config_api.pricing_capability` a few files over: the time to find out is while reading the doctor, not when a 3 a.m. cron stalls.chimera versionShow the Chimera version.chimera featuresShow optional capabilities and what each needs (a key or a dependency).chimera maturityRender the maturity scorecard: surfaces × coverage-IDs proven by real tests.chimera migrateImport config + skills from another agent; --apply also merges long-term memory.chimera models …Model assignment: tier ladder (weak/mid/top), cost mode, and the multi-vendor catalog.chimera secrets …Keep provider keys in the OS vault instead of a file.
Do the work
chimera chatInteractive multi-turn chat — your terminal right-hand. Requires a key. The conversation is saved after every turn, under ``<home>/sessions``, and picked up again on the next run. It is the same store the desktop app reads, so a thread started here can be continued there and the other way round.chimera sessionsList the saved conversations — the same ones the desktop app shows. One store, two front ends. A thread started in the terminal opens in the app, and a thread started in the app can be resumed here with ``chimera chat -s <id>``.chimera tuiLaunch the full-screen TUI — your right-hand. Requires a key.chimera assistYour daily-driver assistant: cheap by default, escalates when it must. Assist = chat with the second-brain defaults ON: the tier cascade routes chit-chat to cheap models and escalates hard asks; your persistent profile (chimera profile) is the stable preamble; memory, nudges and end-of-session consolidation are active. On exit it prints the session cost receipt — tier distribution + measured tokens — so 'cheap by default' is a number.chimera runRun a single-shot Tier-1 completion (no fusion). Requires a provider key.chimera agentRun the ReAct agent loop with native tools. Requires a provider key.chimera deliverDeliverable Mode: produce a polished, self-contained artifact. Requires a key.chimera solveTier-2: autonomously solve a task with plan + verify-or-revert. Requires a key.chimera solve-batchSolve several tasks concurrently, each in its own git worktree (Tier-3 isolation). Every task runs against an isolated checkout, so parallel edits never collide. On merge-back, a file two tasks both changed is reported as a conflict and left for you to resolve rather than silently overwritten. Needs a git repo to isolate.chimera crewRun a multi-agent crew on a task (Tier 3). Requires a provider key.chimera crew-isolatedTier-3: tool-using workers split ONE task, each in its own git worktree, verify-gated. Define workers with repeated --worker 'name:instruction'. Each runs a real agent loop (search/read/edit) against an isolated checkout; non-conflicting edits that pass --verify merge back, files two workers both changed are flagged as conflicts, and a worker whose check fails is rejected (its edits discarded). Needs a git repo to isolate.chimera lifecycleSDLC crew: plan -> build -> test -> review with verify-or-revert. Requires a key.chimera metaMeta-agent: design a specialized agent blueprint for a task. Requires a key.chimera exploreLocate relevant code via the isolated Context Explorer subagent (FastContext-style). Returns only a compact file:line evidence block — the exploration turns never touch your context. A cheap model is usually the right call here; localization is a narrow task.chimera findSearch a repository by what code DOES, not by the string it contains. `chimera/rag/` has been in the tree since 0.44.0 — symbol-level chunking over Python's AST, one SQLite file with an FTS5 index, RRF fusion — measured, documented, and reachable from nothing. A library with no entrance is a library nobody has. This is the entrance. Keyword by default; `--semantic` fuses it with embeddings, and the fusion is what was measured and adopted in `bench/rag/RESULTS.md`: hybrid 0.5050 against keyword 0.4425 on this repository, +6.25 pp paired over 400 probes, McNemar p = 1.7e-04. **`--semantic` means HYBRID, never vectors alone**, and that is the measurement rather than a preference: the vector arm on its own scored **0.4100 — worse than keyword**. Every point of the win comes from fusing two rankings that are wrong about different things. A flag that gave you the vector arm would be a flag that made your search worse. The recall figure is printed with every search because it is per-corpus and per-embedder: the same harness measures 0.4750 on this repository as it stood three weeks ago, and there is no conversion from one embedding model's vector space to another's.chimera workflowRun a declarative workflow — a designed loop — from a YAML file. Requires a key.chimera driftDrift gate: check the workspace against a spec (Spec Growth). Exit 1 on drift.chimera scenariosRun the daily right-hand scenario suite (live). Requires a key.
Fusion and hierarchy
chimera fuseRun a prompt through the LLM-Fusion engine (panel -> judge -> synthesizer).chimera fusion-receiptsSummarize persisted fusion receipts into an honest cost×quality curve.chimera orchestrateHierarchical run: top model decomposes/synthesizes, budgeted mid workers execute. Write-shaped and trivial tasks FALL BACK to the single-agent path by design (the evidence says multi-agent loses there); the fallback is logged with its counterfactual so `chimera delegations` shows the decision.chimera briefMorning brief: parallel topic research through the hierarchy, one synthesized digest. The recipe IS the decomposition (no top-model decompose call). Delegation receipts land in <home>/delegations.jsonl — `chimera delegations` shows what the brief cost vs the inline counterfactual, measured.chimera delegationsMeasured vs counterfactual across delegations — what the hierarchy actually saved.
Memory, profile and inventory
chimera memory …Curated long-term memory.chimera profile …Persistent user profile — the assistant's stable, cacheable preamble.chimera playbook …ACE strategy playbook — incremental, delta-curated guidance for the agent.chimera skillsList the built-in skills.chimera toolsList the built-in native tools.
Skills and evolution
chimera skills-libraryBrowse the curated skill cards that ship with Chimera. Data, not code: each is a markdown page of Trigger/Do/Avoid/Check/Risk that the agent reads into its prompt when it matches. Load one into your own store with ``chimera skills-import <name>``.chimera skills-pendingList learned skills held for review (e.g. distilled during a tainted run).chimera skills-statsPer-skill usage stats (uses, successes, win rate) + retirement candidates.chimera skills-approveApprove/reactivate a learned skill after review (activates retrieval). Works for both a pending skill (held from a tainted run) and a retired one (un-retire).chimera skills-exportExport a learned skill to the open SKILL.md format (portable to the agent-skills ecosystem).chimera skills-importImport a SKILL.md into the store. A tainted-provenance skill is held pending for review. Accepts the plain name of a curated card (``chimera skills-import verify-before-claiming``) as well as a path. The documented form was ``skills/<name>``, a repo-relative path that resolves only inside a checkout — so the one line the README gives for using the shipped library failed for everybody who installed Chimera instead of cloning it. Validated on the way in. This is the only path by which a skill written by somebody else enters the store, and it was the only one that skipped the validator the agent's own proposals must pass — the gate was applied to the code we wrote and not to the code we were handed, which is backwards. A skill card ends up in the system prompt, so an unvalidated one is an instruction from a stranger with the standing of an instruction from the owner.chimera skills-retirePropose retiring under-performing skills — review-gated, never a delete. Retiring only flips status to 'retired' (excluded from retrieval, still inspectable and reactivatable with ``skills-approve``). With no name, acts on the ``retirement_candidates`` signal (used often, low win rate). Dry-run by default; pass ``--apply`` to commit.chimera skills-lifecycleRun the measured skill-lifecycle loop (M18-4): promote proven provisionals, demote regressions. Decisions come from the store's MEASURED usage stats (never a model's self-report): a provisional skill that earns a high win rate over enough uses is promoted to active; a provisional that fails probation or an active skill whose win rate regresses is retired (kept for review). Dry-run by default; ``--apply`` closes the loop — cron it for a hands-off promote/demote cycle.chimera skills-evolveReflectively evolve a skill's prompt template against graded instances (GEPA). Each instance is a `{input, expect}` pair; the (simple, honest) scorer gives 1.0 when the produced output contains the `expect` substring, else 0.0. GEPA reflects on a failing case to rewrite the template and keeps a Pareto frontier of candidates. Dry-run by default: the improved skill is only written back to the store with ``--apply``, and only if it beats the seed.chimera skills-catalogBrowse the installable skills from the wider Agent Skills ecosystem. These are other people's skills, fetched from their repositories on request — not bundled here. The table says what each one NEEDS, because most were written for a different harness and a catalogue that hid that would be advertising features that fail after the download.chimera skills-installDownload a skill bundle from its source repository into your skills directory. Fetches; runs nothing. The bundle lands **pending**: its files are on disk and no part of it reaches a prompt until you approve it. That is the same rule an imported card follows — a skill from a stranger has the standing of an instruction from the owner — and a bundle is that plus executable scripts, so it holds with more reason, not less.chimera skills-bundlesList the skill bundles installed on this machine, and where each came from.chimera skills-bundle-enableSwitch an installed bundle on, so the agent may use it. Do this after reading it. An enabled bundle's name and description reach the agent's prompt when they match a task, and its instructions can tell the agent to run the scripts that came with it — which is why nothing is on by default.chimera skills-bundle-disableSwitch a bundle off, keeping it on disk. Off is not uninstalled, deliberately: trying several and leaving two running is the normal way to use these, and making "off" mean "delete" would charge a download for every change of mind. Use ``skills-uninstall`` when you want the files gone.chimera skills-uninstallDelete an installed skill bundle and its files.chimera evolve …Opt-in model evolution (curate trajectories -> LoRA/DPO recipe).
Automation and projects
chimera cron …Manage scheduled jobs (crons and event SOPs).chimera kanban …Task board with worker lanes (backlog/doing/review/done).chimera project …Run a project start-to-finish against a Spec (drift = acceptance authority).chimera agents …The agents you dispatch work to — as distinct from the one you converse with.
Serve and connect
chimera serveRun the messaging gateway on HTTP, Discord, Telegram, Slack or Signal. Requires a key. Add ``--cron`` to also fire scheduled jobs on a real clock — turning the reactive gateway into an agent that acts on time (the daemon that makes proactivity real). Pass ``--mcp`` to instead expose Chimera *as* an MCP server on stdio, so any MCP client (Claude Desktop, an IDE, another agent) can call ``chimera_solve`` / ``chimera_fuse`` / ``chimera_memory_search``.chimera appRun the Chimera Desktop app: the HTTP+SSE API + the built React UI (needs the 'desktop' extra). Serves the same-origin SPA (``apps/desktop/dist``) and a streaming chat API over the real agent stack. Install with ``pip install 'chimera-agent[desktop]'`` and build the UI once with ``npm --prefix apps/desktop run build``.chimera mcp …Configure MCP servers (persisted to .chimera/mcp.json). Terminal-first source of truth.chimera a2a-cardPrint Chimera's A2A Agent Card JSON (serve it at /.well-known/agent.json).chimera acpServe Chimera to an editor over the Agent Client Protocol (stdio). The mirror of what `chimera code --provider claude` does: there we drive somebody else's agent, here somebody else's editor drives ours. Point Zed, JetBrains or Neovim at `chimera acp` and the loop, the verifier and the receipt are available without installing a second tool. Nothing on this path may write to stdout — it IS the protocol. A stray print corrupts the frame the editor is parsing, and the symptom is an editor that hangs rather than output in the wrong place. The banner goes to stderr for the same reason the MCP server's does.
Safety
chimera guardShow the governance verdict (allow/warn/review/block) for an action.chimera redteamRed-team the injection defenses: attack success rate with vs without them. No key needed — measures whether the governance layer blocks a harmful tool call once a run is tainted (defense-in-depth coverage), not model susceptibility.chimera approveAnswer a decision the agent is waiting on, from anywhere. Without a terminal the approval gate collapsed to a refusal: `ask` degraded to `deny`, so every REVIEW verdict on the VPS, in a container or under cron was a no, and the mandate that says "confirm before billing, before a destructive migration, before touching RLS" had nothing to confirm with. The question is written down and sent to wherever this deployment delivers; this is how it gets answered. Silence is still a refusal — a question times out. That is deliberate: a gate that reads silence as consent produces a record of an approval nobody gave.
Benchmarks and evaluation
chimera benchRun the continuous-evolution benchmark on a demo task set. Requires a key.chimera measure …Run the rulers this project measures itself with.chimera bench-compareReport the honest A/B delta (+95% CI) between two benchmark result files. Feed it the pass/fail from two runs on the SAME task IDs (e.g. a terminal-bench free-model baseline vs the same model driven by Chimera). Prints each arm's Wilson-bounded pass rate, the delta, its Newcombe CI, and whether the difference is significant. This is the number that proves (or doesn't) that the scaffolding lifts a weak model. With --paired, the two lists are treated as *aligned pairs* (each index is one task replayed from an identical forked checkpoint), and the tighter McNemar/Wilson interval is reported — the payoff of running both arms from the same forked state.chimera swe-bench-compareHonest A/B over two SWE-bench Verified-Mini reports on the SAME instance ids. Reads the official evaluation reports (``resolved_ids`` or a per-instance map) for a free model alone vs the same model driven by Chimera, projects both onto the shared instance list (a missing id counts as unresolved), and prints the delta + 95% CI. This is the second standard scoreboard for the weak-model-lift thesis; the pass/fail comes from SWE-bench's tests, never self-reported.chimera fusion-benchA/B the fusion engine: full vs selective (tokens + accuracy). Calls real models.chimera cascade-benchFour-arm bench: weak-only vs mid-only vs cascade vs fusion. Calls real models. Published criterion (stated up front): cascade >= mid-only pass rate at materially lower tokens-per-pass. The number reported is whatever is measured.chimera hierarchy-benchPaired A/B: single-agent (all docs inline) vs the hierarchy (one worker per doc). Calls real models. Both arms run on the SAME model so the comparison isolates the ORCHESTRATION (minimal-context scoping + budgets + contracts), not model strength. Quality = paired McNemar/Wilson (the only place "significant" appears); tokens = measured totals per arm, with no significance claim on cost. `--multistep` switches to the companion suite where the token crossover lives: a single agent re-sends every document on every turn (cost grows with turns), while scoped workers pay each doc ~once — and prices the measured cache reduction via the caching model.chimera skillcard-benchA/B reasoning with vs without injected TRS skill cards. Calls real models.chimera schema-benchMeasure tool-schema token cost, full vs compacted (advertise-time). No model calls.chimera sandbox-benchState-based bench: grade the final workspace state + count harmful side effects. Unlike the text benches, this measures what the agent DID (files it changed), and flags mutations outside each task's allowed set. Uses real models + file tools.chimera memory-benchMeasure recall@k as memory grows — lexical vs paraphrase. Default (keyword search, no key needed) surfaces the honest ceiling: exact-token recall holds at scale, but paraphrase recall collapses. Pass ``--semantic`` to re-run with the embedding recall path and watch the paraphrase column lift — that delta is the whole point of M11b.chimera memory-poisonAblate the memory-poisoning defenses: what reaches a LATER run's prompt, and unmarked. No key needed, nothing leaves the machine. `redteam` measures one run — content arrives untrusted, the harmful call is refused, and the whole picture ends with the process. This measures the other shape: run A stores what it "learned" from a poisoned page, run B asks an unrelated question days later, and recall hands the planted fact to the model. The headline is what arrives **unmarked**, not what is blocked. A poisoned fact carrying its origin is one the model was warned about; an unlabelled one is indistinguishable from something the agent verified itself. Each of the three layers (taint / gate / label) is switched off in turn, because a single number would be compatible with any of them doing nothing. See `bench/memory_poison/PREREGISTRATION.md` for the thresholds, fixed before the first run.chimera probe-selectPROBE best-arm identification with a cheap-proxy control variate (M18-5). "Which model/config is best?" where each expensive reward (a real grade) is paired with a cheap proxy (a weak judge) of unknown correlation. PROBE uses the proxy as a control variate so the estimate needs FEWER expensive draws the better the proxy correlates — and stays unbiased when the proxy is useless. Prints each arm's adjusted mean ± interval, the winner, and — if not yet confident — the arm to sample next. Feed it recorded (proxy, reward) observations from a bench.chimera transfer-gatePromote a learned change only if it helps its tuned slice AND doesn't regress a holdout. Guards against *negative transfer* — a GEPA prompt / ACE delta / distilled skill that raises the pass rate on the tasks it was tuned against but REGRESSES on other tasks sharing the capability. Feed the tuned slice's paired pass/fail (baseline vs candidate) and, ideally, a disjoint same-capability holdout's; the verdict is PROMOTE / BLOCK with the paired evidence (exit 1 on BLOCK, for CI). Without a holdout it promotes on the tuned gain alone but flags that transfer was NOT measured.chimera evoclawStress-test continuous-evolution degradation: naive vs guarded. Requires a key.chimera rubric-gradeGrade an answer against an authorable rubric — weighted criteria with a required-criterion veto. Produces a per-criterion breakdown, a single weighted score, and a pass/fail verdict. A required criterion that falls below the gate vetoes the outcome regardless of the weighted score.chimera context-curveDid runs carrying more context do worse? Measured on THIS machine's own logs. Answers with "not enough data" until the pre-registered floors are met — see `bench/context_curve/PREREGISTRATION.md`, which fixed those floors before any data existed.