AIDOCS

Return current MCP server metrics (token usage, tool calls, output guard stats).

Return metrics in Prometheus text exposition format for /metrics scraping.

Search the official MCP server registry. Returns matching servers with install commands.

Get details for a specific MCP server from the registry.

Show circuit breaker states for all tracked MCP servers.

Manually reset a circuit breaker for an MCP server.

Clear pending 📣 run-done notifications.

Check context budget for a session — journal size, estimated tokens, recommendations.

Compact session context — extract key decisions, prune old journal entries. Resets token counters (new context window).

List recent file edits for rollback. Optionally filter by file or session.

Forensic grep across edit history for old/new content.

Summary of all files modified in this session — who edited what, how many times.

Self-introspection: list all registered MCP tools.

Compact session-state badge for editor status bars.

Build semantic search index from code files. Embeds file contents for meaning-based search.

Check semantic search index status — model availability, indexed files/chunks.

Rollback a specific edit — restore the file to its state before the edit.

Unified session-lifecycle tool — one tool, ten modes (Empire directive 2026-05-12).

Bind THIS host session to an AIDOCS-enabled project — the local mirror of the outer gate's project_select.

Clear both reconnect flags in one idempotent call.

Battlefield briefing for a task — anti-reinvention pre-check.

Empire gate-message CRUD — backs the dashboard editor.

Mark a workflow action as completed with evidence. Call after doing the required work.

Basic git operations. op: status, log, diff, add, commit, push, pull, branch, stash.

Start a persistent long-lived conductor agent for a session.

Send a message/command to the running conductor agent.

Stop the running conductor agent and release session claim.

Unified conductor seat operations — one tool, four modes (Empire directive 2026-05-12).

Unified Q&A channel — one tool, five modes (Empire directive 2026-05-12).

Failure-stewardship disposition AND DIAGNOSIS surface — the agent-callable CONSUMER half of the failure ledger (the Stop hook is the producer).

Who does each identity channel say is calling, RIGHT NOW, on THIS call?

For WHAT KIND of block, WHAT HAPPENS. A read-only diagnostic added 2026-08-25 after a plain CSS append (``printf >> x.css``) froze a session under ``run_destructive`` and no surface could say why, or what any other refusal would have cost.

Which AIDOCS build is this, and what is newest deployed / released?

Read-only runtime process-audit ledger (backlog #335 Phase 1) — what subprocesses did this server spawn, why, and when. Rows are recorded by shell_egress_service.audited_popen (spawn + reap); the fingerprint ENFORCEMENT gate (LEGACY_SUBPROCESS_FINGERPRINTS) is a separate, untouched authority — this …

One inter-agent channel for role messages and actor-routed XAACP.

Role-based audit of the CONNECTED agents on the selected project -- the actual interactive agents (conductors), keyed by host_session_id / agent identity, NOT lane subagents. Each shows its messagerie role (conductor/co_conductor/king), bound work session, liveness (its MCP process), agent_memory_ep…

Auto-resolve a scope expansion request if no lane conflict exists. Approves and expands scope automatically, or flags conflict for manual resolution.

Resolve the best host + model + think_mode for a task type.

Get recent conductor agent output (stdout/stderr). Dashboard polls this for live view.

Clear token usage audit data (AUDEL-gated: host-session operator-auth + admin + reason + audit-of-audit; subagents/unauthenticated refused. No token argument — auth comes from the host-session binding only).

Clear tool-call audit events (AUDEL-gated; host-session auth only).

Prune old execution audit events (AUDEL-gated; host-session auth only). keep_days deletes by age, max_events caps total count.

Get token/tool usage broken down by host and agent identity.

Unified lane control + inspection — one tool for everything an agent does with lanes (Empire directive 2026-05-28: process-shaped work lives under its natural parent, no `ai_debug` catch-all).

Unified plan + lane orchestration. 13 ai_plan_* siblings → 6 modes (operator directive 2026-05-28: aggressive consolidation over rename-only).

Worker management — status, list, kill, and resume for plan-task workers. Resolves the `ai_kill` (worker_id) vs `ai_run(action= 'kill')` (run_id) name collision that the operator caught 2026-05-28: workers and shell processes are different subsystems, one parent each.

Walk a session's execution_events Merkle chain and report whether it's intact.

Return every execution_event stamped with this task_id.

Fallback line-range read. PREFER indexed tools FIRST.

Read a byte range of any file as text. Use for non-indexed text files (logs, CSVs, .csproj, .resx, plain configs) or when reading files too large for the line-based tools. For PDFs/Excel/docx, use the structured-parser tools. Soft cap 512KB per call; hard cap 8MB. For larger files, paginate via offs…

Extract text/tables from a PDF. mode: 'text' | 'text_and_tables'. pages: "1-5,8" or "" for all (max 50). Requires the 'office' extra.

Inspect an Excel workbook (read-only). Modes: 'outline' (sheets+headers), 'sheet' (rows, max 500, needs sheet), 'formulas' (needs sheet), 'trace' (parse formula at sheet!cell). Requires the 'office' extra.

Extract paragraphs and tables from a .docx in document order. 'sections' is an optional "1-3" range limiting output to the first N Heading-1 sections. Requires the 'office' extra.

Inspect or query a SQLite file (read-only). Modes: 'tables' (names+row counts), 'schema' (CREATE stmts, optionally filtered by table), 'query' (SELECT-only, capped at `limit`).

Stream a JSONL file with field-level filter + projection.

Create a new file at a relative path with exact content.

Delete a single project-relative file by moving it to the .TRASH/ recovery area, or restore a trashed file (#385). Two-phase confirm with the path in the phrase, so the operator sees exactly which file before agreeing. Single file only; no glob/batch/ directory. Regenerable artifacts (build/cache di…

A file's IDENTITY and EXISTENCE in one tool — create | rename | delete | restore (#958).

Insert content before a specific line. Clearer than ai_replace(mode='lines') insert mode.

Apply multiple edits atomically across one or more files.

Unified replace: mode='string' (old/new_string; old_string capped by edit.str_replace_max_old_chars, default 1000), 'anchor' (start_anchor + replacement + end_anchor span; inclusive=True consumes both anchors instead of keeping them), 'symbol' (index-resolved symbol body rewrite — requires `symbol` …

Return the release-profile config edit policy visible to agents.

DO NOT TOUCH file protection — writes a sentinel header into the file AND records the protecting identity in the SQL registry (the registry, not the header, is authoritative; only the same user or admin+ can remove).

Find files by name or summary. Use modified_since to filter by recency: 'today', '1h', '24h', '7d'. Replaces Glob — returns ranked results with language and role info.

Full-text search across all indexed files. Replaces Grep — returns matches with line numbers. Use | or OR for multi-term. Set regex=true for patterns. Use `glob` to scope the search to a subset of files. Set expand=True to broaden a single-word query via NLP lemma + semantic synonyms (e.g. 'protect'…

Move a code block from source to target file. Atomic: extracts lines, places in target, removes from source. Use for refactoring large files into modules.

Find start and end line of a symbol using the index. Use before extract_block to avoid manual line counting.

Before extracting a block, show what imports and helpers it depends on that won't come with it.

Move a symbol (function/class/method) from source to target file by name. No line numbers needed — uses the index to find boundaries.

Show the largest symbols in a file that are good extraction candidates. Use to plan deslopification.

Full refactor pipeline: find symbol → extract to target → reindex both → detect stale references + dead code. Returns extraction result plus cleanup suggestions.

Read-only slop FINDER (split 2026-05-24: all mutating refactor modes moved to ai_deslop_apply — a guarded Tier-M path). This tool NEVER mutates and issues no write grants.

Guarded APPLY path for deslop refactors — the Tier-M counterpart of the read-only ai_slop scanner.

Return lightweight dependency edges for one indexed code file.

Return the outline (symbols + kinds + line numbers) of a single indexed file. Much cheaper than reading the whole file when you just need to know what's in it — useful before ai_get_symbol_snippet to pick the right symbol, or to locate a container class for a partial definition.

Return an exact code snippet for an indexed outline symbol.

Get symbol info without reading files. Kind: signature, signatures, constructor, constructors, enum, properties, api.

Investigate a concept by ranking container symbols (classes, structs, records, interfaces) that match it. Returns findings plus suggested next tools, with paths pre-granted for follow-up reads.

Unified find tool — replaces all code_find_* and code_search_* tools.

Unified trace tool.

Unified bundle tool — replaces all code_get_*_bundle tools.

Unified schema tool — replaces all schema_find_*, schema_get_*, schema_trace_* tools.

Initialize AIDOCS structure on a new project — creates .MEMORY/, AGENTS.md/CLAUDE.md, and templates.

Ensure the target project has a .mcp.json with the aidocs MCP server entry for Claude Code.

Run strict session-era structural check on a project.

Run legacy-compatible structural check on a project.

Run safe deterministic structural fixes on a project.

Inspect whether legacy runtime files/folders are still present.

Refresh all derived indexes for a project in one call.

Return a consolidated status view for memory, code, and schema indexes.

Return git remote/origin context, including private/public split hints.

Return the active built-in + project-local language descriptor registry summary.

Validate built-in and project-local TOML language descriptors.

Return the available built-in descriptor semantic families/tags.

Show which descriptor would classify a given project-relative path.

Return current MCP capability index status for a project.

Return indexed MCP capability definitions, optionally filtered by query.

Trigger a remote AIDOCS crown deploy of `ref`. Authority (super_admin + session + ref + non-empty reason + consumable confirm) is enforced at the gate; this verifies the AIDOCS_PRIVATE binding + enqueues for the daemon. The owner recorded is the gate-resolved principal. `confirm_token` is consumed b…

Read the status + log of a deploy enqueued by ai_deploy (queued|running|ok|failed).

Immutable issue filing. Modes: file | list.