subcortex

A local decision layer for coding-agent TUIs. A small daemon keeps a typed-decision model warm and answers System-1 questions — is this request simple? is this tool output still needed for it? — in milliseconds, so the big model doesn't spend tokens on them. Hooks, plugins and MCP wire it into 29 terminal coding agents.

Two decision backends:

• jev — hosted typed-decision API (TypeSafe jev-latest, or any endpoint speaking the same wire shape: OpenRouter, Vercel AI Gateway). ~250 ms per decision over a kept-alive connection, $0.042 per million input tokens.
• laya — local model (laya-mlx on Apple Silicon, laya elsewhere). Private, offline, ~10 ms per question. Weaker judgments, so it acts more cautiously (see How good are the decisions?).

What it does inside a TUI

1. Prompt hint — a request the model rates simple gets a one-line, factual note ("the most direct, minimal change is likely sufficient").
2. Output trimming — a large, successful shell output that the model judges not needed for the user's current request is cut to its head and tail; lines in the middle that mention the request or warn about something are kept, and anything that looks like a failure is never touched. The user's latest request is the evidence: without one, nothing is trimmed. Laya doesn't trim unless you opt in.
3. Compaction snapshot — before context compaction, the last few messages are saved to disk (never vetoing anything).
4. Compaction restore — after compaction they are handed back as context, exactly once, and only deleted after they were delivered.

Each TUI gets whichever of these its extension seam can safely support:

TUI	Seam	Hint	Trim	Compaction
Claude Code	hooks	✅	✅ updatedToolOutput	✅ PreCompact → SessionStart
Codex CLI ≥ 0.133	hooks (/hooks trust)	✅	✅ continue:false + reason	✅
Open Interpreter	hooks (Codex engine)	✅	✅	✅
Qoder CLI	hooks	✅	✅	✅
CodeBuddy Code	hooks	✅	— hook gets no output	✅ on auto-compaction
GitHub Copilot CLI ≥ 1.0.67	hooks	✅	✅ modifiedResult	✅ restored with next prompt
Factory Droid	hooks	✅ (not one-shot exec)	— can only append	✅
Qwen Code ≥ 0.16	hooks	✅	— truncates natively	✅
Gemini CLI ≥ 0.27	hooks	✅	— truncates natively	◐ on /compress
Cursor CLI	hooks	✅ interactive	— Shell can't be replaced	✅ restored with next prompt
Kimi Code CLI ≥ 0.33	hooks	✅	— output hook ignored	✅ restored with next prompt
OpenHands CLI ≥ 1.12	hooks	✅	— truncates natively	✅ via its event log
Grok Build	hooks	— output discarded	✅ tagged updatedToolOutput	✅ restored on next shell call
Docker Agent ≥ 1.137	hooks	✅	✅ updated_tool_response	✅ restored with next prompt
Mistral Vibe ≥ 2.25.5	hooks	— no event	✅ post_tool reason (from a turn's 2nd call)	— no event
Letta Code	hooks	✅ interactive	—	—
Junie CLI	hooks	✅ interactive	—	—
Devin CLI	hooks	✅	—	—
OpenCode ≥ 1.1.62	plugin	✅	✅ bash only	✅ into the compaction prompt
Kilo Code CLI	plugin	✅	✅	✅
Amp	plugin	✅	✅	✅ rolling snapshot
Pi ≥ 0.87	plugin	✅	✅	✅ re-inserted after the summary
Cline CLI ≥ 3.0.62	plugin	✅	✅	✅ on the next request
Crush, Goose, Warp, Zed, Kiro, Auggie	MCP	on-demand tools	—	—
Aider	none	—	subcortex wrap for test/lint	—

"—" means the TUI has no seam that can do it without blocking or failing a tool call, so subcortex doesn't try. Per-TUI details, config paths and caveats: docs/tuis.md.

Verified end to end — each real TUI binary, sandboxed, against a mock model API, asserting on what it actually sent to the model: Claude Code 2.1.278, Codex 0.155.1, Open Interpreter 0.0.45, Gemini CLI 0.60.0, Qwen Code 0.24.3, Kimi Code 2.0.2, OpenCode 1.18.31, Grok Build 1.0.40, Pi 0.87.0, Cline 3.0.62, Docker Agent 1.142.0, Mistral Vibe 2.25.5, Factory Droid 0.224.0, Copilot CLI 1.0.87, OpenHands 1.16.0, Letta Code 0.32.15, CodeBuddy 2.156.0, Junie 26.9.21, Devin CLI 3000.11.1, Cursor CLI 2026.09.18, Amp (plugin runtime), and the MCP integrations of Crush 0.96.1, Goose 1.51.0 and Auggie 0.36.0. Qoder and Kiro need a vendor login to run at all; their hook/MCP formats were checked against their shipped code. Settings for MCP servers belong in config.json: some TUIs (Auggie) start MCP servers without your environment variables.

Install

curl -fsSL https://raw.githubusercontent.com/pavlealeksic/subcortex/main/install.sh | sh

That installs the CLI into its own environment (uv, pipx or a private venv) and starts subcortex setup, an interactive walk-through:

1. Backend — Laya (local; installs laya-mlx/laya into a dedicated environment and downloads the model) or Jev (hosted; your API key is kept in the environment or in a mode-600 file).
2. Behaviors — prompt hints, output trimming, compaction snapshots.
3. TUIs — a checklist of every supported TUI, with the ones found on this machine preselected.
4. Review — each file that will change, with its diff on request; every install self-tests its hook commands before writing.
5. Daemon — start it now, and optionally at login (launchd / systemd).

Arrow keys and space in a terminal; plain numbered prompts elsewhere. Run it again any time to change your choices; subcortex setup --yes --backend jev --tuis detected runs it unattended. Or install by hand:

pip install subcortex            # daemon + jev backend (zero dependencies)
pip install "subcortex[laya]"    # + local laya backend

Wire it into a TUI (without the wizard)

subcortex install                         # checklist of TUIs (in a terminal)
subcortex tuis                            # supported TUIs, what's installed, what's on PATH
subcortex install claude-code --dry-run   # show exactly what would change
subcortex install claude-code             # shows the diff, asks, self-tests, writes
subcortex install codex gemini-cli --yes  # several at once, no prompt
subcortex install detected                # every supported TUI found on PATH
subcortex install cursor --mcp            # also register the on-demand MCP server
subcortex uninstall claude-code           # removes exactly what subcortex added

How good are the decisions?

A probability is not an accuracy, so every rule was chosen on labeled examples and then checked on examples it had never seen (docs/calibration.md). The mistakes that matter are a "simple" hint on complex work and a trim of output the request needs; there were none:

	hints: simple requests	hints: complex requests	trims: disposable output	trims: needed output
jev (calibration / held out)	17/20 · 12/12	0/20 · 0/12	12/12 · 6/6	0/12 · 0/6
laya (calibration / held out)	5/20 · 5/12	0/20 · 0/12	off by default¹	—

¹ Laya's output judgments trimmed 2 of 6 needed outputs on the held-out set, so Laya only trims if you set thresholds.output_needed_threshold yourself.

subcortex eval                  # the same check against your backend and thresholds

It exits non-zero on any hinted complex request or trimmed needed output.

Safety

A hook that misbehaves can take a TUI down: an earlier version of this project exited 2 from a Claude Code hook, and exit 2 means "block this prompt". People run this inside real work, so every one of these is enforced by tests:

• Hooks always exit 0 and never stall a turn. One hardened entry point catches everything; a watchdog ends it silently when its budget (4 s) is spent — including a C-level backstop for work that holds the interpreter lock; stray prints never reach stdout; nothing is written to stderr.
• Blocking responses can't be emitted. A guard drops decision: block/deny, continue: false, permission denials and the like — except where a TUI's spec uses one as a non-blocking replacement, for that event only.
• Your environment can't break it. Hooks run as '<python>' -I -m subcortex.hook <tui> <event> 2>/dev/null || true: isolated from PYTHONPATH and from the project directory (a repo's own json.py can't run inside a hook or the daemon), and a missing installation stays harmless.
• Nothing is sent to a proxy. Hooks and plugins reach the daemon over a direct socket to 127.0.0.1, whatever HTTP_PROXY says.
• Many sessions at once are fine. Per-session state is keyed by TUI and session, written atomically, kept private (0600/0700) and changed under a lock; a compaction snapshot is restored exactly once and deleted only after delivery. The daemon answers bursts (backlog 256), sheds load it can't serve in time (503, the hook passes through), and serializes the local model.
• Only you can use the daemon. Every request carries a per-user token from a 0600 file; requests from web pages (Origin, foreign Host, non-JSON) are refused.
• Installs are self-tested. Before writing anything, the installer runs the exact commands it's about to write — with the daemon down, against a stub daemon, and with the executable missing — and refuses to write if any run exits non-zero, writes to stderr, emits a blocking response, runs over budget, or stays silent where it must answer.
• Config edits are conservative. A diff is shown and confirmed first; writes are atomic with timestamped backups; a symlinked (dotfiles) config stays a symlink; a file the TUI changed during install is re-merged, not overwritten; files that aren't plain JSON / valid TOML are refused; only subcortex's own entries are ever removed.
• Other TUIs' hooks are respected. Cursor, Droid, Grok Build, Devin and Cortex Code also execute hooks from ~/.claude/settings.json; the Claude Code adapter recognises foreign payloads and stays silent.

Privacy

• laya keeps everything on your machine.
• jev receives, per decision, your latest request, the tool call and a ~1.5 KB head+tail excerpt of a large output — capped, and with anything that looks like a secret (API keys, tokens, passwords, private keys, credentials in URLs) masked first.
• Local state (remembered requests, compaction snapshots) is private to your user and pruned automatically; the ledger behind subcortex stats holds counts only, never content.

Daemon & settings

subcortex doctor                 # check config, backend, daemon, installed hooks
subcortex config                 # show every setting (config edit: interactive)
subcortex config set thresholds.min_output_chars 8000
subcortex service install        # start the daemon at login (launchd / systemd --user)
subcortex serve                  # start the background daemon (127.0.0.1:7707)
subcortex stats                  # hints, trims (tokens saved), restores and jev cost, per TUI
subcortex eval                   # labeled examples through your backend and thresholds
subcortex serve --stop
subcortex decide --state "Refactor the parser to be iterative" \
    --questions '{"hard": {"type": "noul", "instructions": "Is `prompt` hard?"}}'

HTTP surface (all on 127.0.0.1:<port>; every request but /health carries the per-user token from ~/.local/share/subcortex/token):

• POST /decide {state, questions} → {success, answers, model?, usage?}
• POST /verdict/prompt {prompt} → {success, verdict: {label, confidence, signals}}
• POST /verdict/output {output, context, task} → {success, verdict: {needed, p_needed, signals}}
• POST /v1/prompt-hint, /v1/tool-output, /v1/snapshot, /v1/restore — the four behaviors, for plugins (see docs/adding-a-tui.md)
• GET /health, GET /stats

Configuration

~/.config/subcortex/config.json (SUBCORTEX_CONFIG overrides the path; SUBCORTEX_* env vars override keys):

{
  "backend": "laya",
  "port": 7707,
  "features": {"prompt_hint": true, "trim_output": true, "compaction_snapshot": true},
  "thresholds": {"prompt_simple_confidence": null, "output_needed_threshold": null, "min_output_chars": 6000},
  "hooks": {"autostart_daemon": true, "budget_s": 4.0, "http_timeout_s": 3.0, "head_chars": 1000, "tail_chars": 500,
            "snapshot_messages": 5, "snapshot_chars": 500}
}

null thresholds use the rule calibrated for the active backend; a number replaces that rule's primary threshold (check the effect with subcortex eval). The daemon picks up config changes without a restart.

Hooks start the daemon themselves when they find it down (at most once a minute; set hooks.autostart_daemon to false or SUBCORTEX_AUTOSTART=0 to turn that off) — the hook that notices still passes through untouched. subcortex doctor also checks every installed integration's hook executable.

Runtime state (daemon log/PID/token, compaction snapshots, the stats ledger, hooks.log) lives in ~/.local/share/subcortex/ (SUBCORTEX_DATA_DIR). Set SUBCORTEX_DEBUG=1 to log every hook invocation to hooks.log.

Any other TUI

subcortex mcp     # stdio MCP server: subcortex_decide / _classify_prompt / _judge_output
subcortex wrap -- <command>   # run a command, trim its disposable output, keep its exit code

MCP tools are model-invoked (the agent spends tokens calling them), so hooks or plugins are preferred wherever a TUI has them. Adding a native adapter: docs/adding-a-tui.md.

Development

python3 -m venv .venv && .venv/bin/pip install -e .
.venv/bin/python -m unittest discover -s tests   # no model or API key needed; bun runs the plugin tests
SUBCORTEX_E2E=1 .venv/bin/python -m unittest discover -s tests -p "test_e2e*.py"   # real TUIs, mock LLM API

The unit suite includes concurrency proofs (racing threads and processes, bursts against the daemon), watchdog and proxy-immunity checks, and runs every TS plugin under Bun. The end-to-end suites drive real TUI binaries against a local fake LLM API (tests/e2e/) — no network, no cost — and assert on what each TUI actually sent to the model. They never touch your real TUI configs.

Releases publish to PyPI via Trusted Publishing (.github/workflows/publish.yml): bump version in pyproject.toml, then gh release create v<X.Y.Z> --generate-notes.