Execute a ChimeraLang program. Returns emitted values, confidence scores, gate logs, and assertion results.

Assert value confidence >= 0.95. Returns ConfidentValue on pass, ConfidenceViolation on fail.

Wrap a value as Explore<> — marks it unverified. Use for hypotheses and brainstorms.

Collapse multiple candidates into one consensus result. Strategies: majority, weighted_vote, highest_confidence. Returns winner and divergence score.

Hallucination and MCP security detection. Strategies: range, dictionary, semantic, cross_reference, temporal, confidence_threshold. Returns hallucination flags plus prompt-injection/tool-poisoning signals.

Apply constraint middleware to a tool output. Checks confidence, forbidden capabilities, hallucinations. Returns pass/fail with audit trace.

Static type-check a ChimeraLang program without executing it. Validates confidence boundaries and scope rules.

Execute ChimeraLang and generate a Merkle-chain integrity proof. Returns results + tamper-evident hash chain with root hash and verdict. Vs. direct reasoning: produces a cryptographic hash chain that makes each reasoning step auditable and tamper-detectable — an LLM cannot self-certify its own reasoning steps this way.

Session constraint audit: total calls, pass/fail counts, avg confidence, warnings, tools used, and material/security metadata.

Extract atomic claims from text or an envelope with claim typing, hedge/abstention tagging, and provenance metadata.

Verify claims against evidence using lexical token-overlap scoring. Returns lexically_supported / lexically_contradicted / lexically_insufficient verdicts — IMPORTANT: verdicts are Jaccard token-overlap, not semantic entailment or NLI. Supplement with chimera_constrain for semantic checks. Vs. direct reasoning: provides explicit per-claim scores, curated verification_gold corpus matching, and prompt-injection attack-pattern detection.

Merge multiple result envelopes into one combined envelope with aggregated confidence and trace history.

List, inspect, or apply reusable constraint policies such as strict_factual, brainstorm, medical_cautious, code_review, mcp_security, prompt_injection_hardened, and research_factcheck.

Inspect persisted result envelopes. Actions: latest, get, list, stats. Includes material pack and security metadata.

Inspect bundled material packs and manifests. Actions: list_packs, status, licenses, source_manifest. Read-only and offline.

Full pipeline: optimize documents → compress messages → budget gate. Returns quality_passed, budget_remaining, lossy_dropped_count.

Remove filler, deduplicate sentences, normalize whitespace from text. Returns optimized text and token savings.

Compress text via contractions/symbols. Levels: light, medium, aggressive. Returns compressed text and compression ratio. Token savings are automatically recorded to chimera_dashboard (set auto_track=false to disable). Prefer over manual regex replacement — the quantum algorithm uses structural salience so important sentences near the task focus survive compression.

Token usage vs context window. Returns status (ok/warn/critical) and recommendation (ok/compress/fracture).

Score messages 0–1 for context-window management. mode='drop_priority' (default): scores by recency+type+density — lowest scores are dropped first in lossy compression. mode='importance_for_goal': scores by alignment with the focus goal — highest scores are most relevant to keep. Vs. direct reasoning: O(n) tokenisation is far cheaper than asking the model to rank N messages, which consumes O(N*content) prompt tokens.

Estimate token count and USD cost for text or messages. No API call. Supports Claude, GPT, Gemini models.

Log before/after token counts to the session cost tracker. View with chimera_dashboard.

Session cost summary: tokens saved, dollars saved, avg compression %, last 10 events.

CALL FIRST on every message. Optimizes input, estimates token cost, proposes budget. Show proposal_text to user for approval. After approval: constrain response to max_output_tokens and use optimized_prompt as effective input.

Lock approved token budget after user approves CSM proposal. Actions: lock/check/update/release. Returns remaining_tokens and status (ok/warn/critical). At critical status, compress draft before sending.

Causal graph. Actions: add_edge, query, paths, info.

Multi-perspective deliberation. Default mode 'semantic' uses stance detection + prompt-term alignment + concept overlap — reports consensus_detected:true when >=60% of perspectives share a stance AND avg_similarity>=0.62. Mode 'lexical_consensus' uses raw Jaccard token overlap (faster, but misses paraphrases — use only when vocabulary is controlled). Vs. direct reasoning: externalises the perspective set so callers can inject viewpoints not all present in one model pass, and provides a numeric divergence score the model cannot compute on its own without a separate summarisation step.

Calibration error (ECE) for [{predicted_confidence, was_correct}]. Returns overconfidence/underconfidence rates.

Record adaptation events (context, action, outcome). Actions: record, stats.

Confidence-weighted consensus vote across agent responses. Returns winner, confidence, contradiction count.

Decompose a goal into ordered sub-goals. Detects goal type and returns task list with strategy.

Confidence-weighted key-value world state. Actions: update, query.

Pattern-based content safety check plus MCP security attack-pattern detection. Returns safety verdict, reason, attack flags, and category counts.

Rule-based ethical scoring of an action. Checks non-maleficence, autonomy, justice, beneficence. Returns score and recommendation.

Sensor/action state simulator. Actions: perceive, act, status, reset. Tracks position, perception, action_log, energy.

Interaction history tracker. Actions: record_interaction, query, list_agents. Tracks sentiment and relationship_strength per agent.

Domain analogy mapper. Actions: add_mapping, query, list. Maps concepts from source_domain to target_domain.

Fitness-ranked candidate selector. Runs N generations of selection+mutation. Returns ranked survivors and best candidate.

Capability tracker. Actions: update (record capability+evidence), reflect (return all capabilities).

Keyword-search knowledge base. Actions: add, search, list.

Session memory store with importance scoring. Actions: store, recall.

Returns the relevant tool subset for a task type. Call to avoid unnecessary tool invocations and reduce token overhead.

Execute multiple chimera tools in one call. Saves tokens vs separate round-trips. Returns array of results in call order.

LLM-free extractive summarizer. Ranks sentences by TF-IDF and returns top N. Token savings are automatically recorded to chimera_dashboard (set auto_track=false to disable). Use before passing long docs to other tools.

Build Anthropic prompt-cache markers for stable text blocks (system prompt, CLAUDE.md, tool defs, fixtures). Returns blocks ready to drop into the SDK system array with cache_control: ephemeral on blocks that meet the model's minimum cacheable size. Lossless — beats any compression on bytes that recur every turn (75-90% off cached tokens).

Compress build/test/install logs while preserving every error, warning, and traceback line verbatim. Keeps head + tail windows for context. Typical reduction 80-95% on noisy logs with zero loss of diagnostic signal.

Estimate per-turn baseline cost (system prompt + tool definitions + MCP server registrations). Surfaces the 'ghost tokens' the model pays on every turn so you can prune unused MCP servers or trim verbose tool defs.

Inspect or query the per-namespace tool-call dedup cache. Use action='get' with key (sha256 prefix) to retrieve a prior call's metadata, action='list' to see all tracked calls, action='clear' to reset. Populated automatically by the PostToolUse hook.

End-of-session summary: total tokens saved, dedup cache hits, top compressed responses, lock state. Safe to call any time; the Stop hook calls it automatically when the agent stops responding.

Emit a system instruction that forces an AI agent to write only in Chimera Glyph (CG) — a compact AI-only pidgin designed for token efficiency. The returned directive should be injected as a system message; the calling AI then produces CG output, and the human-visible translation happens via chimera_glyph_translate.

Translate Chimera Glyph (CG) text back into readable English. Lossy by design: reconstructs meaning, not surface form. Use this at the end of an AI-internal CG reasoning chain to produce a human-readable summary.