abs

Aggregated admin snapshot (billing/security/compliance/beta/vault).

Claude Haiku 4.5 — fast Anthropic model. Short tasks, classification.

Claude Sonnet 4.6 — balanced quality/speed. Default for code and analysis.

Claude Opus 4.7 — strongest Anthropic model. Deep analysis, critical work.

GPT-OSS 20B (Groq) — ultra fast (<0.4s). Short questions, classification.

Llama 4 Scout 17B (Cloudflare) — instruction following, short tasks.

Cerebras — 235B MoE, ~0.3s latency, 1M tokens/day on the free tier.

Gemini 2.5 Flash — fast multimodal. Templates, short generation.

Gemini 2.5 Pro — 1M context, deep analysis, multimodal.

CloudFlare Llama 3.3 70B FP8 Fast — edge latency.

CloudFlare GPT-OSS 120B — 120B at the edge; alternative to the Groq route.

Kimi K2.5 (CloudFlare) — code generation and planning. 256K context.

Phi-4 via local Ollama — reasoning. Requires OLLAMA_URL to be set.

Beta waitlist counts, recent signups, paid-conversion rate.

Estimated cost so far today: tool usage priced against provider configs.

The most recent N learnings, plus per-category counts.

Record a learning by hand. category: bugfix|delegation|arch|security|perf|ux.

Billing dashboard: Stripe products, license counts, revenue, recent webhooks.

What this install can do with the keys it has, and what one more buys.

Ask a subscription CLI whether it can actually answer.

Index a workspace root into the call-graph (run before querying it).

What breaks if `target` (symbol name or file path) changes — its transitive callers.

Direct callers of a symbol.

Symbols related to `symbol` within N hops (callers + callees, undirected).

Cohere usage, latest alert, and severity (ok|warn|danger|limit_hit).

The most recent N alerts, newest first.

Mark an alert acknowledged.

Cohere Command R+ 08-2024 — enterprise chat, RAG-friendly.

Cohere Aya Expanse 32B — 101 languages; the multilingual workhorse.

Cohere embed-english-v3.0 — returns a 1024-dim embedding as JSON.

The approval queue for this tenant: items plus the risk-bucket summary the Approval Center header shows. `status`: pending | approved | rejected | all.

Approve, reject or edit one pending approval item. `decision`: approve | reject | edit. An item that was already decided is refused, not rewritten.

Recorded meetings for this tenant, newest first: status, duration, speaker count, summary and whether the transcript reached the knowledge base.

One meeting: summary, speakers, the transcript segments, and the action items detected in it — what a task panel turns into work.

The workflows this tenant has saved: name, how many steps, when it was last touched. Reading only — running one is not exposed here, because a workflow can call out to the network and spend money, and that decision belongs behind the approval path rather than a panel button.

The schedules that actually fire for this account: when each next runs, when it last ran, and the reason it stopped if it did.

Schedule a saved workflow: 5-field cron, UTC (e.g. '0 9 * * 1' = Mondays at 09:00). Scheduled runs use the free provider chain only. An expression this scheduler cannot run is refused with the reason, rather than stored as a schedule that silently never fires.

Stop a schedule. Only this account's own schedules are removable here.

GDPR compliance posture dashboard.

Propose a multi-file edit for `task`, graded + blast-annotated, applied to nothing.

Groq Compound — multi-step agentic (planning + tool calling).

Groq Compound Mini — fast agentic (short multi-step tasks).

Demo readiness snapshot (demo mode flag, seed version, assets).

Onboarding email queue dashboard.

Live per-provider request quota for this tenant: used, left, percent, throttle state and reason. Providers with no known limit are reported as unlimited rather than shown as an empty gauge.

Tell the server which project the caller has open.

Ask the provider cascade and return the answer WITH its provenance: which provider answered, the failover trail, tokens, latency and estimated cost. Use this instead of ask_* when the caller shows the user where the answer came from.

Run a multi-model pipeline and return its STEPS as structured data — step name, model, elapsed, ok — so the caller can draw the delegation chain instead of parsing a summary line. `kind`: qual_code | qual_analysis | qual_tr | race | race_code.

Which providers this account has its own key for. Metadata only — plaintext keys are never returned, by this tool or any other.

Store this account's own key for a provider (BYOK), encrypted at rest. The key is scoped to the CALLING USER — organisation-wide keys stay in the panel, so a delegated editor token cannot rewrite an org's credentials. The value is never echoed back.

Remove this account's own key for a provider. Only the caller's own key is removable here — the org's key is not this token's to delete.

The cascade in the order it will actually be tried for THIS caller: position, free or paid, whose key answers, breaker and throttle state.

Everything the editor's title bar shows, in one call.

Text to insert at the cursor — the Tab / ghost-text completion.

Detect the layer of a prompt: frontend|backend|database|devops|testing|docs|architecture.

Layer-aware code generation. layer="auto" detects the layer from the prompt.

Inventory a project directory: file counts by extension, size, manifests.

Scan a project, then have a model turn the inventory into a task plan.

Gemini 2.5 Flash Image — generate an image from a prompt (base64 PNG).

Gemini Image Pro (Nano Banana Pro) — higher-quality image generation.

Edit a base64 image according to the prompt.

Start a Veo 3.0 video job. Returns an operation name — video generation is long-running, so poll it with gemini_video_status.

Poll a video job by the operation name gemini_video returned.

Block until a video job finishes, polling every 15s, up to max_seconds.

Gemini Flash Lite — cheapest, fastest single-shot answer.

Ask a question about the contents of a URL (Gemini fetches it itself).

Answer grounded in Google Search, with the source URLs appended.

Schema-guaranteed JSON output. schema_json must be a valid JSON schema.

Arm freeze mode: Write/Edit is allowed only under project_dir.

Arm investigate mode: hooks warn on an edit that was not preceded by reading the code. An empty topic disarms it.

Symbol DB substring search — name LIKE %q%, opsiyonel kind=function|class|import.

RAG hybrid retrieval — BM25 + cosine fusion. alpha_semantic 0=BM25, 1=cosine.

predict the acceptance probability of these scores with an ML model.

Judgment averages over the last N days: drift_signal, outcome_counts, top_files.

The most recent N judgments (id, ts, file, ast/llm/combined, outcome).

Record the outcome of a judgment (accept|reject) — this is the training signal.

Current persona thresholds, last training run, and history size.

Retune the persona from recorded judge outcomes. Below min_samples it refuses with 'insufficient_data' rather than fitting to noise.

Reset the persona to its defaults. The outcome history is kept, so a later train() can rebuild from the same evidence.

Licence and demo status, as JSON.

Demo countdown durum (started/expired/days_remaining).

5-source AI/dev news digest (Anthropic / OpenAI / Gemini / GH / MCP).

Create or update a note. Pass note_id to update an existing one.

List notes, newest first (title + preview).

Fetch one note in full by id.

Lexical search over notes.

Delete a note by id.

Results of the last benchmark run (cascade, vault, symbol, watchdog).

QUALITY CODE — generate (kimi + gpt-oss-20b in parallel) -> verify -> fix.

QUALITY TURKISH — generate (qwen32b + gemini in parallel) -> review -> polish.

QUALITY ANALYSIS — three independent takes (gptoss, kimi2, gemini-pro) -> synthesis.

QUALITY TRANSLATION — translate -> back-translate -> compare -> refine.

RACE — gpt-oss-120b vs kimi vs kimi2 in parallel; first success wins.

RACE CODE — CF Kimi K2.5 vs Groq GPT-OSS 120B; first success wins.

RACE TR — Qwen32B vs Gemini 2.5 Flash; first success wins.

RACE LOCAL — Ollama phi4 vs gemma2. Requires ABS_OLLAMA_URL.

QUAL + HUMANIZE — rewrites the qual-tr output to read less like model output.

QUAL CODE + HUMANIZE — rewrites the qual-code output without the model's narration comments.

Heuristic 'AI-written' score for a text (0 = human, 1 = model). Returns JSON.

Verify code with three local models in parallel — security, tests, lint.