List supported games with their currently indexed build and patch version.

This server's DB identity: db_path, data_dir, pid, started_at and version, plus status.

    Compare it against the other server's `GET /api/health` before writing so two servers pointed
    at different data dirs (a split store) are caught instead of silently diverging.

Full-text search across everything. Each hit's key feeds the detail tool for its kind: convar/command→get_convar(name), entity→get_entity('side:classname'), schema_class/schema_enum→ get_schema_class('module:Name'), event→get_game_event, script→get_script_api, pulse→search_pulse, example→get_example(key), doc→get_doc(key). With game set, general docs are included too.

Get a console variable or command: default, min/max, flags (cheat, replicated, ...) and help text.

List console variables/commands, filtered by substring and/or flag.

Get a Hammer entity: keyvalues (with choices), inputs and outputs, including inherited ones. defined_in shows which base declares each member.

Find entity classes (including base classes) that declare an input or output with this name. Subclasses inherit it too, so check candidates with get_entity.

Get an engine schema class or enum: fields with types, defaults and metadata, the inheritance chain and direct subclasses. Use it for networked fields, VData keys (Deadlock abilities) and plugin offsets.

List schema classes/enums by name substring, module, kind or declared field.

Get a game event's fields and types (the effective definition after core/game/mod overrides).

List game events, optionally filtered by a substring of the name or description.

Look up the CS2 cs_script (point_script) JavaScript API. A type name returns the type with all members, including inherited ones. Anything else returns the matching members with signature and docs. Instance is the Domain type.

List every top-level cs_script type: the Instance const, classes, interfaces, enums and type aliases.

Search Pulse (Valve's visual scripting) node bindings: entity inputs/outputs and functions.

Show what changed between two ingested builds, at member granularity: one row per added/ removed/changed schema field, entity member, game-event field, script symbol, convar, Pulse binding or Hammer class. Each entry carries group (the parent class/entity/event/library) so members cluster under it; before/after hold only that member's own value. counts gives the exact per-section added/removed/changed breakdown ({section: {added, removed, changed}}), taken BEFORE the per-section limit cap, so it stays true even when sections is truncated to the first limit entries; totals[section] is the sum of that section's three counts. Use it after a game patch to find code or configs that might break.

List example scripts (Valve's cs_script_demo samples) with the API each one uses.

Get the full source of an example script. Prefer adapting these Valve-written patterns.

Get a community documentation page (source2.wiki) as Markdown, with its URL and license for citation.

Browse community documentation pages (titles, URLs) without bodies.

Type-check a cs_script with the TypeScript compiler against this build's Valve point_script.d.ts. Errors with category unknown_api mean the method/property does not exist — fix before shipping. needs_type_narrowing gives the subclass to cast to. ALWAYS run this before handing a script to the user.

Check every command in a .cfg against the indexed convars/commands: unknown names (with suggestions), out-of-range values, cheat/devonly flags, and client-vs-server context.

Create a new addon in the local workspace (content/csgo_addons// with scripts/, maps/, materials/, ...), a starter cs_script, Valve's point_script.d.ts + tsconfig.json, and a README with the Hammer setup steps. Fails if the addon already exists.

Write a Source 2 model document (.vmdl) for a mesh that is already inside the addon.

    `mesh` is addon-relative (e.g. models/probe/probe.dmx). The document copies the shape the
    game's own compiler accepts — compile it on a host with resourcecompiler afterwards
    (build_start). Returns the written path. Use overwrite=true to replace an existing .vmdl.

List addons in the local workspace.

Show an addon's local root and file list.

Read a text file from an addon in the workspace.

Write a text source file (.js, .cfg, .vmat, .vdata, panorama .xml/.css, ...) into an addon. Binary/compiled assets are produced by Valve's tools, not written here. Use overwrite=true to replace.

List orchestration workflows: multi-step recipes (research → design → build → validate → test → document) that chain this server's tools and prompts. Start one with start_project.

Start a persistent orchestration project and get the first ready step(s). Progress is stored on the server, so any later session can continue with next_step(project_id).

Get the ready step(s) of a project: instructions, tools to use, rendered prompt text and the gate the server will check when the step is marked done.

Record progress on a step and get the next brief. Marking 'done' runs the step's gate (e.g. all addon scripts must type-check); if it fails you get the problems back — fix them and retry.

Full project state: every step's status, notes, artifacts and recent events.

List orchestration projects (most recently updated first).

Render one of this server's prompts as text, for clients that can't fetch MCP prompts mid-task.

List registered test/build hosts with their facts (shell, GPU, paths) and current lock state. CHECK THIS before driving any host: a fresh lock means another session/workstream owns it. Each lock carries phase (what the owner is doing), phase_age_sec and stuck — the SERVER marks a lock STUCK when its heartbeat is fresh but the phase has not moved past the configured threshold (liveness is not progress). Treat lock.stuck=true as "the owner may be hung".

One host's registry entry: ssh target, shell flavor, probed facts, and lock.

Register (or update) a test/build host. Upsert: omitted fields keep their previous values.

SSH to the host, run the probe script (GPU, RAM, processes, CS2 path/build) and store the facts. Use before planning work on a host — e.g. to see if a game or a heavy GPU process is running.

Claim exclusive use of a host. Fails if another owner holds a fresh lock — a stale lock (heartbeat older than its TTL, e.g. from a crashed worker) is taken over automatically. Always report a phase (what you are doing); the server flags a lock STUCK when its heartbeat is fresh but the phase has not moved for too long, so keep the phase current (pass a new phase to heartbeat_host) as you progress. Heartbeat while you work; release with release_host when done.

Refresh your lock's heartbeat so it doesn't go stale mid-task. Heartbeating the SAME phase does not clear stuck-ness (that is the point); report a new phase as your work moves on.

Release a host lock. Always call this when finished with a host.

Start a server-owned build run on a host: acquires the host lease, runs the command over SSH and returns a run id immediately. Do NOT poll with sleeps — use build_wait(run_id) or build_status. The host lease is released automatically when the run reaches a terminal status. The lease's phase is the run kind (e.g. 'compile'), so a run whose compile hangs while its heartbeat stays fresh reads STUCK on list_hosts — the server, not the client, decides that.

One build run's record: status, exit code, parsed summary counts, material guard and artifacts.

    `warnings` carries the MATERIAL GUARD findings when the log referenced missing/reserved materials
    (Valve's `_vrf/` subtree, or another map's `maps/<other>/` subtree): counts per kind plus the first
    few offending paths. Such a run can still be 'succeeded' — the resourcecompiler exits 0 and the map
    then error-renders as a red/black checkerboard, so treat a non-None `warnings` as NOT clean.

Tail a build run's log. The log is UTF-16 (resourcecompiler) — this decodes it for you, returning the detected encoding, whether the tail was truncated, and any MATERIAL GUARD warnings found in the full log (see build_status).

Block (server-side, no client sleep loop) until the run finishes, then return its record. Raises if it is still running after timeout_sec. With strict_materials, a run whose log has material-guard warnings is an error even though its status is 'succeeded'.

Cancel a run: mark it cancelled and release its host lease (terminates the process when this server owns it).

Recent build runs (newest first), optionally one host and/or active only.

Copy files to a host and verify what landed (Get-FileHash over SSH). Returns a deploy id whose entries carry the server-checked md5 + verified flags. Call deploy_verify later to re-check.

    The push takes the host lease for the copy+verify window: a host held by another owner is a
    conflict (pass that owner if it is you), a free host is claimed as 'deploy:<id>' and released
    when the copy finishes. Inside an orchestration project the project already holds the lease as
    'project:<project_id>', so pass that as `owner` (the `cs2_live_test` deploy step does).

Re-check a recorded deployment's MD5s on the host and update its status.

Recent deployments (newest first), optionally one host.

Record structured evidence for a step. This is what the evidence gate checks — a step with an evidence gate CANNOT be marked done until at least one non-note row exists for it. Use kind='note' only for a supplementary remark (it does not satisfy the gate).

    Pass `step_id` to attach the row to a project step, or omit it to record evidence at project
    level (a measurement, a console transcript, a negative result). A step-less row satisfies no
    step gate, by design.

List evidence rows (newest first), optionally filtered by project, step and/or kind.

Start a server-side test session (state=running). Claims the host lease when a host is given. Mirror the state onto the in-game TEST STATUS HUD yourself via iso_test start over netcon.

Set the running session's current step (the HUD strip's iso_test step). Running runs only.

Finish a running session pass/fail (the HUD's iso_test pass|fail): releases the host lease and records a console-transcript evidence row on the linked project when there is one.

Dismiss a session (iso_test off): state=idle, releasing a still-held host lease.

One test run's record, or the current active run when no id is given.

Recent test sessions (newest first), optionally filtered by state.

Jobs newest first. status='active' means queued + running. Each row carries error/crash counts.

One job with its full timeline (events), links to test runs/builds/deploys/evidence and artifacts.

A job's screenshots tied to their step and verdict (seq, step, reason, url, web_url, verdict, dimensions) — the image context a reviewer pulls after a run. Filter by step/reason.

Enqueue a job (validated intake -> queued). The intake is the gate: every missing required field comes back in one error with a field list. Set host_choice='ask' to queue unassigned.

The head queued job for a host, only when that host's lease is free. Errors (with the holder) when a fresh lock blocks it, and when the host's queue is empty.

Set (or move) a queued job's host and queue order.

queued -> running, recording the worker and started_at. A second call is a conflict.

Append one timeline event. phase updates the job's current phase; error/crash require a message and SHOULD carry data.log_tail so the job page can expand the failure in a .

Copy a local file into the job's artifacts (MCP runs on the Mac). Append-only: no overwrite. Images also record a screenshot event automatically.

Ask the runner for a screenshot: an info event plus params.pending_shot=true. The runner polls the job and answers by uploading an artifact, which clears the flag.

Attach a canonical record the job is about (a test run, build, deploy, evidence row or project).

Answer the running job's current step with a human verdict (pass | fail | skip). Writes a human-confirm evidence row through the same path as POST /api/evidence (so the confirm push files fire) and appends a milestone event. The job must be running and link a test run, else the call is a conflict.

Close out a running/queued job. Gate: summary must carry what, why, learned and next, each non-empty and <=600 chars; otherwise the call is rejected with the offending field names. status='inconclusive' is the honest close-out for a run that ended without a pass/fail verdict (a step-gated demo the user stopped answering) — it is NOT a pass.

Correct a FINISHED job's terminal status (terminal -> terminal), for a mis-recorded close-out — e.g. a run that finished 'passed' but never got a verdict. reason is mandatory and lands as a milestone timeline event. The job must already be terminal (a queued/running job finishes or is cancelled) and the target must differ from the current status, else the call is a conflict.

queued/running -> cancelled; the reason is recorded as a timeline event.