Skip to content

CHANGELOG.md

Latest commit

 

History

History
98 lines (79 loc) · 4.15 KB

CHANGELOG.md

File metadata and controls

98 lines (79 loc) · 4.15 KB

Changelog

v2.0 — Closing the agent loop

v1 made what your agent did visible. v2 makes that visibility actionable by feeding it back into the agent's reasoning loop.

Highlights

  • Five Claude Code hooks registered automatically by tma1-server install --adapter claude-code: UserPromptSubmit prepends a session digest before each turn, SessionStart orients fresh sessions with prior state, PreCompact carries critical state through context compaction, PostToolUse injects per-tool anomaly notes, and Stop blocks termination on unresolved high-severity issues.
  • Seven MCP stdio tools: get_context_bundle, get_session_state, get_anomalies, get_build_status, get_external_changes, get_project_state, get_peer_sessions. The MCP server is the same tma1-server binary, invoked via mcp-serve.
  • Six anomaly rules running on every Detect, with a 10-minute per-session suppression layer + per-rule resolution checks. Each anomaly routes through a Channel (user_prompt_submit, stop_block) so the same finding never injects twice.
  • /tma1-peer slash command reads what Codex / OpenClaw / Copilot CLI just left on this project and brings it into Claude's context — no copy-paste between terminals.
  • Three new sensors populate the perception bundle:
    • Build sensor — wraps build / dev commands and captures output to tma1_build_events (tma1-server build [--watch] -- <cmd>).
    • Git/file sensor — fsnotify + periodic git log -1 poll, classifies changes as human or agent and writes to tma1_external_changes.
    • Project sensor — language / build-system / structural index, refreshed daily into tma1_project_state.

One-shot adapter install

# macOS / Linux
curl -fsSL https://tma1.ai/install.sh | TMA1_ADAPTER=claude-code bash

# Windows
$env:TMA1_ADAPTER = 'claude-code'; irm https://tma1.ai/install.ps1 | iex

This wires hook entries into ~/.claude/settings.json, registers TMA1 as an MCP server in ~/.claude.json, drops the /tma1-peer skill into ~/.claude/skills/, and adds a TMA1 block to your project's CLAUDE.md (or AGENTS.md). All idempotent — repeat install only updates what's stale.

New HTTP endpoints

Endpoint Use
/api/hooks Hook event ingest (request-response, returns injection content)
/api/hooks/stream SSE feed of hook events for the live agent canvas
/api/anomalies Recent anomalies across sessions
/api/anomalies/budget Daily emit count per Kind (1.7 gate ≤ 5 / day)
/api/anomalies/follow-rate Did the agent take the suggested action within N tool calls? (1.7 gate ≥ 30%)

New GreptimeDB tables

Table Source
tma1_build_events Build sensor (tma1-server build)
tma1_external_changes Git/file sensor
tma1_project_state Project sensor
tma1_anomaly_emits Suppression layer logs every emitted anomaly

The 27 hook event types (tma1_hook_events) and conversation table (tma1_messages) from v1 are unchanged. tool_input derived columns (tool_file_path, tool_command_prefix, tool_success, tool_error_summary) are extracted at ingest so anomaly rules don't have to regexp_match the raw blob.

Configuration additions

Variable Default Effect
TMA1_ADAPTER (empty) Wire an agent during install (claude-code). Idempotent.
TMA1_DISABLE_INJECTION (unset) When 1, hook handlers return empty bodies.
TMA1_CONTEXT_PRESSURE_THRESHOLD 100000 Token threshold for context_pressure anomaly.

Compatibility

Every v1 dashboard view, every v1 endpoint, every v1 table, and the entire OTel proxy path are unchanged. v2 is additive: if you don't install an adapter, TMA1 still works as the v1 local-first observability tool with no behaviour change.

Documentation