One name per role: src/gcontext moves to gcontext/ (flat layout), the
instructions files split into framework-instructions.md (G0) and the
project's agent.md (G1). Adds the built-in describe-first setup prompt,
state files as MCP resources (gcontext://<path>), and --version.
Exec: run_script splits into run_script and run_adhoc_script; results are
plain readable text with a status line, output_schema=None on all six
tools kills the {"result": ...} wrapper. 100k-char stream caps.
Writes: connection.yaml is agent-writable (only secrets.env stays
blocked), write_file returns a unified diff and warns on index.md drift,
and the instructions require user approval before every write. New
learn-from-errors rule records fail-then-fix lessons where they belong.
Fixes: secrets.load() strips surrounding quotes (quoted tokens caused
misleading 401s). Tests: 61 passing.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
State changes (from the maat-agent spikes):
- run_script runs saved scripts by path with args and params (PARAM_<NAME>
env vars), plus ad-hoc code; results start with a status header and a
missing-package hint points at connection.yaml deps
- commands/ folders in connections and modules register as MCP prompts
(<owner>__<command>), surfaced as slash commands; .md and .py formats
- instructions.md is pushed at connect through the MCP handshake and
declared as ledger pipe G0: one controllable file is what the agent
receives at start
- read_context/write_context renamed to read_file/write_file; new list_dir
and grep tools
- stateless HTTP: server restarts no longer strand attached clients
Removed:
- flows (never met a real use case; modules plus commands cover process
needs; design.md records the return condition)
- gcontext chat (redundant once the handshake delivers instructions); the
controlled claude invocation is documented in the README instead
- docs/templates (duplicated README sections)
- the ledger's dual chat/mcp mode, collapsed to one
Structure:
- server.py is only the MCP surface; concerns split into fs.py, exec.py,
secrets.py, state.py, ledger.py, commands.py; agent-facing tool text
lives in prompts/tools/*.md
- read-only web dashboard served at the root: overview, ledger, files,
live activity feed (web/ Vite app, bundled into the wheel)
- secrets.env is now unreadable through the agent (read guard)
39 tests. Version 0.4.0.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
When neither --port nor gcontext.yaml pin a port and the default is taken,
up picks the next free one and persists it to gcontext.yaml so the URL stays
stable. init no longer ships a toy connection; the README and init output now
walk through adding a real first connection instead.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Local-first state manager for AI agents. One HTTP server (gcontext up),
any MCP runtime attaches by URL. Context ledger (nothing reaches the agent
invisibly), secrets isolation (names visible, values injected and scrubbed),
flows (multi-step work computed purely from files), archive convention.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
