apunkt/iai-mcp-opencode
1
0
Fork 0
Code Issues Pull requests 3 Projects Releases 1 Packages Wiki Activity Actions

Compare commits

base: apunkt:main
Branches Tags
apunkt:main
apunkt:renovate/tsx-4.x-lockfile
apunkt:renovate/typescript-7.x
apunkt:renovate/node-26.x
apunkt:opencode/issue8-20260522085810
apunkt:oc-1.0
..
compare: apunkt:opencode/issue8-20260522085810
Branches Tags
apunkt:renovate/tsx-4.x-lockfile
apunkt:main
apunkt:renovate/typescript-7.x
apunkt:renovate/node-26.x
apunkt:opencode/issue8-20260522085810
apunkt:oc-1.0

No commits in common. "main" and "opencode/issue8-20260522085810" have entirely different histories.
main ... opencode/i

16 changed files with 94 additions and 1408 deletions
Download patch file Download diff file Expand all files Collapse all files

33
README.md
View file

@ -133,23 +133,15 @@ Use the absolute path. `~` and `$HOME` won't expand here.
For Claude Desktop (untested), edit `~/Library/Application Support/Claude/claude_desktop_config.json`.
### Install the OpenCode plugins
### Install the OpenCode plugin
**Capture plugin** — listens for idle sessions and writes a deferred-capture JSONL file for the daemon to drain.
This is the OpenCode equivalent of the Claude Code Stop hook. It listens for idle sessions and writes a deferred-capture JSONL file for the daemon to drain.
```bash
mkdir -p ~/.config/opencode/plugins
cp deploy/opencode/iai-mcp-capture.js ~/.config/opencode/plugins/
```
**Memory-inject plugin** — automatically loads session background memory on every new session via the daemon's HTTP adapter. Memory is injected into the model's system prompt (clean, no phantom messages, preserves session title).
```bash
cp deploy/opencode/iai-mcp-memory-inject.js ~/.config/opencode/plugins/
```
This plugin requires the daemon's HTTP listener. Set `IAI_DAEMON_HTTP_PORT=0` in your daemon's systemd/launchd config and restart it. The daemon writes the live port to `~/.iai-mcp/.http.port`, which the plugin reads automatically.
Make sure `@opencode-ai/plugin` is installed:
```bash
@ -157,27 +149,12 @@ cd ~/.config/opencode
npm install @opencode-ai/plugin
```
Register both plugins in `~/.config/opencode/config.json`:
```json
{
"plugin": [
"~/.config/opencode/plugins/iai-mcp-capture.js",
"~/.config/opencode/plugins/iai-mcp-memory-inject.js"
]
}
```
### Connect OpenCode
Add the MCP server and plugins to `~/.config/opencode/config.json`:
Add the MCP server to `~/.config/opencode/config.json`:
```json
{
"plugin": [
"~/.config/opencode/plugins/iai-mcp-capture.js",
"~/.config/opencode/plugins/iai-mcp-memory-inject.js"
],
"mcp": {
"iai-mcp": {
"type": "local",
@ -188,7 +165,7 @@ Add the MCP server and plugins to `~/.config/opencode/config.json`:
}
```
Use the absolute path to `mcp-wrapper/dist/index.js`. Both plugin files must already be in `~/.config/opencode/plugins/`.
Use the absolute path to `mcp-wrapper/dist/index.js`. The plugin file (`iai-mcp-capture.js`) must already be in `~/.config/opencode/plugins/`.
### Verify
@ -385,7 +362,7 @@ Limitations worth knowing about:
Claude Code is the primary host, validated in daily use.
OpenCode is supported via the `iai-mcp-capture.js` and `iai-mcp-memory-inject.js` plugins (see Install the OpenCode plugins above) and MCP server config in `~/.config/opencode/config.json`. The memory-inject plugin requires the daemon's HTTP listener (`IAI_DAEMON_HTTP_PORT` env var).
OpenCode is supported via the `iai-mcp-capture.js` plugin (see Install the OpenCode plugin above) and MCP server config in `~/.config/opencode/config.json`.
Claude Desktop should work (uses `claude_desktop_config.json` instead of `~/.claude.json`) but hasn't been tested end to end.

147
deploy/opencode/iai-mcp-memory-inject.js
View file

@ -1,147 +0,0 @@
/**
* iai-mcp memory-injection plugin for OpenCode (approach A).
*
* Pulls session-start memory from the iai-mcp daemon's localhost HTTP adapter
* and injects it directly into the model's system prompt via the
* `experimental.chat.system.transform` hook. The memory CONTENT is placed in
* context — no tool call, no injected "INIT" user turn, so the session title
* is generated from the user's real first message (clean).
*
* REPLACES iai-mcp-session-init.js — do NOT run both. session-init forces a
* tool call via a phantom turn (hijacks the title); this plugin needs neither.
*
* Requires the daemon's HTTP listener:
* - set IAI_DAEMON_HTTP_PORT in the daemon's systemd unit (e.g. "0" for an
* OS-assigned port) and restart it. The daemon writes the live port to
* ~/.iai-mcp/.http.port, which this plugin reads.
*
* wake_depth=standard is requested so l0/l1/l2/rich_club carry real content
* (minimal mode returns only opaque handles — nothing worth injecting).
* Override via IAI_MCP_WAKE_DEPTH (minimal|standard|deep).
*
* Fail-safe: any error is swallowed; system-prompt assembly must never break.
*/
const HOME = process.env.HOME || process.cwd();
const PORT_FILE = `${HOME}/.iai-mcp/.http.port`;
const WAKE_DEPTH = process.env.IAI_MCP_WAKE_DEPTH || "standard";
const FETCH_TIMEOUT_MS = 10000; // a cold runtime-graph build can exceed 5s
const MAX_ATTEMPTS = 3; // give up after this many consecutive failed fetches per session
const INJECT_MAX_TURNS = 3; // stop injecting after N turns (token-cost control)
const memo = new Map(); // sessionID -> injected text | null (null = permanent failure)
const inflight = new Map(); // sessionID -> Promise (await for concurrent callers)
const attempts = new Map(); // sessionID -> consecutive-failed-fetch count
const turns = new Map(); // sessionID -> number of times memory was injected
async function readPort() {
const fs = await import("node:fs");
try {
const port = parseInt(fs.readFileSync(PORT_FILE, "utf8").trim(), 10);
return Number.isInteger(port) && port > 0 ? port : null;
} catch {
return null; // daemon HTTP not enabled / not up yet
}
}
async function fetchMemory(sessionId) {
const port = await readPort();
if (!port) return { ok: false, text: "" };
const url =
`http://127.0.0.1:${port}/memory/session-context` +
`?session_id=${encodeURIComponent(sessionId)}` +
`&wake_depth=${encodeURIComponent(WAKE_DEPTH)}&format=text`;
try {
const res = await fetch(url, { signal: AbortSignal.timeout(FETCH_TIMEOUT_MS) });
const text = (await res.text()).trim();
// Distinguish network failure from valid-but-empty response.
// An empty body on HTTP 200 is a valid result (new user, minimal mode).
return { ok: res.ok, text };
} catch {
return { ok: false, text: "" };
}
}
// Centralised fetch+cache so the warm-on-create event and the per-turn
// transform hook share ONE in-flight fetch, one cache, and one attempt budget.
// This dedupes concurrent calls so a session never emits duplicate daemon-side
// session_started events.
//
// Fix: inflight stores a Promise so concurrent callers await the same fetch
// instead of getting "". Fix: memo stores null for permanent failure so we
// don't retry forever. Fix: distinguish {ok:true,text:""} (valid empty) from
// {ok:false} (network error).
async function ensureMemory(sessionId) {
// Already resolved — return cached value (may be "" or null).
const cached = memo.get(sessionId);
if (cached !== undefined) return cached;
// A fetch is already running — await it instead of returning "".
const existing = inflight.get(sessionId);
if (existing) return existing;
// Permanent failure — stop retrying.
if ((attempts.get(sessionId) || 0) >= MAX_ATTEMPTS) {
memo.set(sessionId, null);
return null;
}
const promise = (async () => {
const { ok, text } = await fetchMemory(sessionId);
if (ok) {
// HTTP 200 — memoize regardless of content (empty is valid).
memo.set(sessionId, text);
attempts.delete(sessionId); // reset failure counter on success
} else {
// Network error — increment counter for retry budget.
attempts.set(sessionId, (attempts.get(sessionId) || 0) + 1);
}
return text;
})();
inflight.set(sessionId, promise);
try {
return await promise;
} finally {
inflight.delete(sessionId);
}
}
export const IaiMcpMemoryInject = async () => {
return {
// Warm-on-create: prime the daemon graph cache (and our memo) as soon as a
// session appears — before the user's first turn — so the first transform
// injects immediately instead of paying the cold-build latency.
event: async ({ event }) => {
if (event.type !== "session.updated") return;
const sid = event.properties?.info?.id;
if (!sid) return;
try {
await ensureMemory(sid);
} catch {
// never throw from an event handler
}
},
"experimental.chat.system.transform": async (input, output) => {
try {
const sid = input?.sessionID;
if (!sid || !output || !Array.isArray(output.system)) return;
// Turn-limited injection: stop after INJECT_MAX_TURNS to control token cost.
// Memory was already in the system prompt for the first N turns; the model
// has seen it and doesn't need repetition every turn.
const currentTurns = turns.get(sid) || 0;
if (currentTurns >= INJECT_MAX_TURNS) return;
const text = await ensureMemory(sid);
if (text) {
turns.set(sid, currentTurns + 1);
output.system.push(`# iai-mcp memory (session start)\n${text}`);
}
} catch (err) {
// NEVER throw — a plugin error must not break system-prompt assembly.
console.error(`[iai-mcp] memory inject failed: ${err.message}`);
}
},
};
};

47
deploy/opencode/iai-mcp-session-init.js
View file

@ -1,47 +0,0 @@
/**
* @deprecated Use `iai-mcp-memory-inject.js` instead.
*
* This plugin forces a tool call via a phantom "INIT" user turn, which
* corrupts session title generation. The memory-inject plugin uses the
* system.transform hook to inject memory directly into the system prompt
* — no phantom turn, clean titles, memory visible every turn.
*
* Retained temporarily for backward compatibility. Will be removed in a
* future release. Migrate to iai-mcp-memory-inject.js:
*
* cp deploy/opencode/iai-mcp-memory-inject.js ~/.config/opencode/plugins/
*
* Then replace this plugin reference in ~/.config/opencode/config.json.
*/
export const IaiMcpSessionInit = async ({
client,
project,
directory,
worktree,
experimental_workspace,
serverUrl,
$,
}) => {
const doneSessions = new Set();
return {
event: async ({ event }) => {
if (event.type !== "session.updated") return;
const sid = event.properties?.info?.id;
if (!sid || doneSessions.has(sid)) return;
doneSessions.add(sid);
try {
await client.session.prompt({
path: { id: sid },
body: {
system: "MANDATORY: Call `memory_session_context` (no arguments) BEFORE answering any user query.",
parts: [{ type: "text", text: "INIT" }],
noReply: false,
},
});
} catch {}
},
};
};

33
mcp-wrapper/package-lock.json generated
View file

@ -8,15 +8,16 @@
"name": "iai-mcp-wrapper",
"version": "0.1.0",
"dependencies": {
"@modelcontextprotocol/sdk": "^1.0.0"
"@modelcontextprotocol/sdk": "^1.0.0",
"zod": "^4.0.0"
},
"bin": {
"iai-mcp-wrapper": "dist/index.js"
},
"devDependencies": {
"@types/node": "^25.0.0",
"@types/node": "^22.0.0",
"tsx": "^4.7.0",
"typescript": "^6.0.0"
"typescript": "^5.4.0"
},
"engines": {
"node": ">=18"
@ -517,13 +518,13 @@
}
},
"node_modules/@types/node": {
"version": "25.9.5",
"resolved": "https://registry.npmjs.org/@types/node/-/node-25.9.5.tgz",
"integrity": "sha512-OScDchr2fwuUmWdf4kZ9h7PcJiYDVInhJizG/biAq3cAvqwYktuy/TYGGdZNMtNTFUP7rnb0NU4TUdm82kt4Rg==",
"version": "22.19.17",
"resolved": "https://registry.npmjs.org/@types/node/-/node-22.19.17.tgz",
"integrity": "sha512-wGdMcf+vPYM6jikpS/qhg6WiqSV/OhG+jeeHT/KlVqxYfD40iYJf9/AE1uQxVWFvU7MipKRkRv8NSHiCGgPr8Q==",
"dev": true,
"license": "MIT",
"dependencies": {
"undici-types": ">=7.24.0 <7.24.7"
"undici-types": "~6.21.0"
}
},
"node_modules/accepts": {
@ -1584,9 +1585,9 @@
}
},
"node_modules/tsx": {
"version": "4.23.0",
"resolved": "https://registry.npmjs.org/tsx/-/tsx-4.23.0.tgz",
"integrity": "sha512-eUdUIaCr963q2h5u3+QwvYp0+eqPvn+egeqZUm0hwERCqqx1E3kK5ehbGCvqSE5MQAULr67ww0cA3jKc3YkM1w==",
"version": "4.22.3",
"resolved": "https://registry.npmjs.org/tsx/-/tsx-4.22.3.tgz",
"integrity": "sha512-mdoNxBC/cSQObGGVQ5Bpn5i+yv7j68gk3Nfm3wFjcJg3Z0Mix9jzAFfP12prmm5eVGmDKtp0yyArrs0Q+8gZHg==",
"dev": true,
"license": "MIT",
"dependencies": {
@ -1617,9 +1618,9 @@
}
},
"node_modules/typescript": {
"version": "6.0.3",
"resolved": "https://registry.npmjs.org/typescript/-/typescript-6.0.3.tgz",
"integrity": "sha512-y2TvuxSZPDyQakkFRPZHKFm+KKVqIisdg9/CZwm9ftvKXLP8NRWj38/ODjNbr43SsoXqNuAisEf1GdCxqWcdBw==",
"version": "5.9.3",
"resolved": "https://registry.npmjs.org/typescript/-/typescript-5.9.3.tgz",
"integrity": "sha512-jl1vZzPDinLr9eUt3J/t7V6FgNEw9QjvBPdysz9KfQDD41fQrC2Y4vKQdiaUpFT4bXlb1RHhLpp8wtm6M5TgSw==",
"dev": true,
"license": "Apache-2.0",
"bin": {
@ -1631,9 +1632,9 @@
}
},
"node_modules/undici-types": {
"version": "7.24.6",
"resolved": "https://registry.npmjs.org/undici-types/-/undici-types-7.24.6.tgz",
"integrity": "sha512-WRNW+sJgj5OBN4/0JpHFqtqzhpbnV0GuB+OozA9gCL7a993SmU+1JBZCzLNxYsbMfIeDL+lTsphD5jN5N+n0zg==",
"version": "6.21.0",
"resolved": "https://registry.npmjs.org/undici-types/-/undici-types-6.21.0.tgz",
"integrity": "sha512-iwDZqg0QAGrg9Rav5H4n0M64c3mkR59cJ6wQp+7C4nI0gsmExaedaYLNO44eT4AtBBwjbTiGPMlt2Md0T9H9JQ==",
"dev": true,
"license": "MIT"
},

4
mcp-wrapper/package.json
View file

@ -18,8 +18,8 @@
"@modelcontextprotocol/sdk": "^1.0.0"
},
"devDependencies": {
"@types/node": "^25.0.0",
"typescript": "^6.0.0",
"@types/node": "^22.0.0",
"typescript": "^5.4.0",
"tsx": "^4.7.0"
},
"engines": {

23
mcp-wrapper/src/tools.ts
View file

@ -24,7 +24,6 @@ export const TOOL_NAMES = [
"memory_contradict",
"memory_capture",
"memory_consolidate",
"memory_session_context",
"profile_get_set",
"curiosity_pending",
"schema_list",
@ -186,24 +185,6 @@ export const toolSchemas: Record<ToolName, ToolSchema> = {
},
},
},
memory_session_context: {
name: "memory_session_context",
description:
"Retrieve the current session context payload (identity, session handle, " +
"topic cluster, rich club). Call at session start to load relevant " +
"background memory. Returns JSON with l0, l1, l2, rich_club, and " +
"compact_handle fields. Optional session_id parameter to query a specific session.",
inputSchema: {
type: "object",
properties: {
session_id: {
type: "string",
description:
"Optional session id to query. When omitted, uses the wrapper's current session.",
},
},
},
},
profile_get_set: {
name: "profile_get_set",
description:
@ -367,10 +348,6 @@ export async function invokeTool(
return bridge.call("memory_capture", args);
case "memory_consolidate":
return bridge.call("memory_consolidate", args);
case "memory_session_context": {
const sessionId = (args.session_id as string) || null;
return bridge.call("session_start_payload", sessionId ? { session_id: sessionId } : {});
}
case "profile_get_set": {
const op = args.operation as string;
if (op === "get") {

7
src/iai_mcp/capture.py
View file

@ -201,13 +201,6 @@ def capture_turn(
log.exception("capture_turn insert failed")
return {"status": "skipped", "record_id": None, "reason": f"insert-failed: {type(e).__name__}"}
# Wire temporal_next: link this insert to the previous same-session insert.
try:
from iai_mcp.retrieve import link_temporal_next
link_temporal_next(store, rec, session_id=session_id)
except Exception:
pass # diagnostic only; never block the write path.
return {"status": "inserted", "record_id": str(rec.id), "reason": f"tier={tier}"}

11
src/iai_mcp/core.py
View file

@ -741,15 +741,6 @@ def dispatch(store: MemoryStore, method: str, params: dict) -> dict:
# wake_depth knob reaches the assembler.
from iai_mcp.session import assemble_session_start, SessionStartPayload
sid = params.get("session_id", "-")
# D5-02 per-call wake_depth override: a caller (e.g. the HTTP adapter
# serving session-context for system-prompt injection) can request
# standard/deep content WITHOUT mutating the global profile knob.
# Invalid/absent values fall back to the per-process profile state.
profile_state = _profile_state
wd_override = params.get("wake_depth")
if wd_override in ("minimal", "standard", "deep"):
profile_state = dict(_profile_state or {})
profile_state["wake_depth"] = wd_override
records_count = store.count_rows("records")
if records_count == 0:
empty = SessionStartPayload(
@ -765,7 +756,7 @@ def dispatch(store: MemoryStore, method: str, params: dict) -> dict:
payload = assemble_session_start(
store, assignment, rc,
session_id=sid,
profile_state=profile_state,
profile_state=_profile_state,
)
return _payload_to_json(payload)

29
src/iai_mcp/daemon.py
View file

@ -1403,27 +1403,6 @@ async def main() -> int:
mcp_socket = SocketServer(store, lock=lock, state=state)
mcp_socket_task = asyncio.create_task(mcp_socket.serve())
# Tier-1 PoC: optional localhost HTTP adapter over the same core.dispatch.
# OFF by default -- enabled only when IAI_DAEMON_HTTP_PORT is set (value "0"
# means OS-assigned ephemeral port, written to ~/.iai-mcp/.http.port for
# discovery). Lets MCP hosts that lack a server->client push mechanism
# (e.g. OpenCode) fetch session memory by URL. NO auth in this tier --
# 127.0.0.1 bind is the only boundary; see http_server.py header.
http_server = None
http_task = None
_http_port_env = os.environ.get("IAI_DAEMON_HTTP_PORT")
if _http_port_env is not None:
try:
from iai_mcp.http_server import HttpServer
http_server = HttpServer(store, port=int(_http_port_env))
http_task = asyncio.create_task(http_server.serve())
except (ValueError, OSError):
# Bad port value or bind failure must NOT crash daemon boot
# (mirrors the maintenance try/except above). The unix socket
# and all lifecycle tasks come up regardless.
http_server = None
http_task = None
# Plan 10.6-01 Task 1.4: REMOVED `_propagate_idle_shutdown`
# bridge task. The socket-side `idle_watcher` (which set
# mcp_socket.shutdown_event after IDLE_CHECK_INTERVAL_SECS of
@ -1682,20 +1661,12 @@ async def main() -> int:
mcp_socket.shutdown_event.set()
except Exception:
pass
# Drain the optional HTTP adapter the same way before cancellation.
if http_server is not None:
try:
http_server.shutdown_event.set()
except Exception:
pass
_cancel_targets = [
tick_task, audit_task, s4_task, cascade_task,
mcp_socket_task,
cpu_watchdog_task,
lifecycle_tick_task,
]
if http_task is not None:
_cancel_targets.append(http_task)
for t in _cancel_targets:
t.cancel()
# Drain task exceptions silently: we're shutting down.

19
src/iai_mcp/graph.py
View file

@ -42,7 +42,6 @@ class MemoryGraph:
def __init__(self) -> None:
self._nx: nx.Graph = nx.Graph()
self._ig: "ig.Graph | None" = None
self._ig_stale: bool = False
self._attrs: dict[UUID, dict[str, Any]] = {}
self._backend: str = "networkx"
@ -80,11 +79,10 @@ class MemoryGraph:
self._nx.add_edge(
str(src), str(dst), weight=weight, edge_type=edge_type
)
self._maybe_switch_backend()
if self._backend == "igraph":
# Lazy rebuild: igraph mirror is stale after edge writes
# and will be rebuilt on next read.
self._ig_stale = True
if self._ig is not None:
# igraph mirror is immutable by topology; rebuild after each edge
# write while in igraph backend. Cheap enough at Phase-1 scale.
self._rebuild_igraph()
# ------------------------------------------------------ backend switching
@ -107,17 +105,11 @@ class MemoryGraph:
weights = [
float(self._nx[u][v].get("weight", 1.0)) for u, v in self._nx.edges()
]
g = ig.Graph(n=len(nodes), edges=edges, directed=False) # type: ignore[unbound]
g = ig.Graph(n=len(nodes), edges=edges, directed=False)
g.vs["name"] = nodes
if weights:
g.es["weight"] = weights
self._ig = g
self._ig_stale = False
def _ensure_igraph(self) -> None:
"""Rebuild igraph mirror if stale from deferred edge writes."""
if self._ig_stale and _HAS_IGRAPH:
self._rebuild_igraph()
# ---------------------------------------------------------- graph metrics
@ -132,7 +124,6 @@ class MemoryGraph:
bc = nx.betweenness_centrality(self._nx, weight="weight")
return {UUID(n): float(c) for n, c in bc.items()}
# igraph path
self._ensure_igraph()
assert self._ig is not None
has_weight = "weight" in self._ig.es.attributes()
raw = self._ig.betweenness(weights="weight" if has_weight else None)

326
src/iai_mcp/http_server.py
View file

@ -1,326 +0,0 @@
"""Tier-1 localhost HTTP adapter over core.dispatch (PoC).
Sibling transport to socket_server.py: both are thin adapters around the
single transport-agnostic entry point ``core.dispatch(store, method, params)``
(D7-08 -- one dispatch function per method, no transport branching). This one
speaks HTTP/1.0 on 127.0.0.1 so an MCP host that has *no* server->client push
mechanism (e.g. OpenCode, which ignores the MCP ``instructions`` field) can
still pull session memory into its system prompt by fetching a URL.
PoC SCOPE / NON-GOALS (deliberately minimal -- Tier 1):
- localhost-only bind. NEVER 0.0.0.0. The unix socket's security boundary is
filesystem perms (chmod 0o600, T-04-07); a TCP port has none, so we limit
blast radius to loopback and ship NO auth here. Token auth is the Tier-2
follow-up before this is anything but a local convenience.
- one request per connection (Connection: close). No keep-alive, no chunked
encoding, no HTTP/1.1 pipelining. Simplest robust thing for fetch() clients.
- disabled unless IAI_DAEMON_HTTP_PORT is set (see daemon.main wiring). Absent
=> no TCP surface at all, identical to today.
Constitutional guards (mirror socket_server.py):
- C-DISPATCHER-FSM-ISOLATION: this adapter calls core.dispatch ONLY; it never
touches the daemon FSM. FSM transitions stay owned by daemon.py's tick.
- C3 ZERO API COST: stdlib + core.dispatch only; no SDK references.
- C5 LITERAL PRESERVATION: transport-only; zero record mutation here.
- R3 head-of-line: dispatch is sync + 50-500 ms, so it runs on
asyncio.to_thread exactly like socket_server.py:257.
- R5 fail-loud: dispatch raises map to HTTP status codes (see _STATUS_FOR).
"""
from __future__ import annotations
import asyncio
import json
import os
import sys
import time
from pathlib import Path
from typing import Any
from urllib.parse import parse_qs, urlsplit
from iai_mcp.concurrency import SOCKET_PATH
from iai_mcp.core import UnknownMethodError, dispatch
# Discovery file: consumers (OpenCode plugin) read the live port from here,
# mirroring the SOCKET_PATH convention (~/.iai-mcp/.daemon.sock). Sits next to
# the socket so both transports share the one well-known dir.
HTTP_PORT_FILE: Path = SOCKET_PATH.parent / ".http.port"
# dispatch-exception -> HTTP status, mirroring socket_server.py's JSON-RPC code
# mapping (ERR_METHOD_NOT_FOUND/-32601, ERR_INVALID_PARAMS/-32602, ...).
_BAD_REQUEST = 400
_NOT_FOUND = 404
_METHOD_NOT_ALLOWED = 405
_INTERNAL = 500
_MAX_BODY_BYTES = 1 << 20 # 1 MiB cap on POST bodies; refuse larger with 400.
def _payload_to_text(result: Any) -> str:
"""Flatten a session_start_payload dict into a system-prompt-ready block.
Best-effort and forgiving: any missing/empty key is skipped, so this is
correct at every wake_depth:
- standard/deep populate the human-readable content layers
(l0/l1/rich_club/l2) -- the useful case for system-prompt injection.
- minimal (the default) leaves those empty and emits only the compact
pointer/handle fields (~30 tok). Those are opaque references, not
content; we still surface them on a trailing line so the block is
never silently empty, but a host that wants real memory text in the
prompt must run wake_depth>=standard.
Token-accounting ints and breakpoint_marker are dropped.
"""
if not isinstance(result, dict):
return str(result)
parts: list[str] = []
headers = {"l0": "## L0 Identity", "l1": "## L1 Recent Summary", "rich_club": "## Rich Club"}
for key in ("l0", "l1", "rich_club"):
val = result.get(key)
if isinstance(val, str) and val.strip():
parts.append(f"{headers[key]}\n{val.strip()}")
l2 = result.get("l2")
if isinstance(l2, list):
for item in l2:
text = item if isinstance(item, str) else json.dumps(item, ensure_ascii=False)
if text.strip():
parts.append(f"## L2 Episode\n{text.strip()}")
# Compact handles (populated at minimal; also present at standard/deep).
handles = [
result.get(k)
for k in ("identity_pointer", "compact_handle", "brain_handle", "topic_cluster_hint")
]
handles = [h for h in handles if isinstance(h, str) and h.strip()]
if handles:
parts.append(f"## Handles\n{' '.join(handles)}")
return "\n\n".join(parts)
class HttpServer:
"""Per-connection HTTP/1.0 server routing to core.dispatch.
Routes (all GET unless noted):
- GET /healthz -> {"ok": true} (no dispatch)
- GET /memory/session-context -> dispatch("session_start_payload",
{session_id?}); ?format=text returns
a flattened text/plain block.
- POST /rpc {method, params} -> dispatch(method, params); generic
passthrough mirroring the socket.
Constructor args mirror SocketServer: a shared MemoryStore singleton plus
bind knobs. shutdown_event lets daemon.main drain gracefully.
"""
def __init__(
self,
store: Any,
*,
host: str = "127.0.0.1",
port: int = 0,
port_file: Path | None = None,
) -> None:
self.store = store
self.host = host
self.port = port # 0 => OS-assigned ephemeral port (written to port_file)
self.port_file = port_file if port_file is not None else HTTP_PORT_FILE
self.bound_port: int | None = None
self.last_activity_ts: float = time.monotonic()
self.active_connections: int = 0
self.shutdown_event: asyncio.Event = asyncio.Event()
# -- request parsing ----------------------------------------------------
async def _read_request(
self, reader: asyncio.StreamReader
) -> tuple[str, str, dict[str, str], bytes] | None:
"""Parse one HTTP request. Returns (method, target, headers, body) or None on EOF."""
request_line = await reader.readline()
if not request_line:
return None
try:
method, target, _version = request_line.decode("latin-1").rstrip("\r\n").split(" ", 2)
except ValueError:
raise _HttpError(_BAD_REQUEST, "malformed request line")
headers: dict[str, str] = {}
while True:
line = await reader.readline()
if line in (b"\r\n", b"\n", b""):
break
raw = line.decode("latin-1").rstrip("\r\n")
if ":" in raw:
name, _, value = raw.partition(":")
headers[name.strip().lower()] = value.strip()
body = b""
length = headers.get("content-length")
if length is not None:
try:
n = int(length)
except ValueError:
raise _HttpError(_BAD_REQUEST, "bad content-length")
if n > _MAX_BODY_BYTES:
raise _HttpError(_BAD_REQUEST, "body too large")
if n > 0:
body = await reader.readexactly(n)
return method.upper(), target, headers, body
# -- routing ------------------------------------------------------------
async def _route(self, method: str, target: str, body: bytes) -> tuple[int, str, str]:
"""Return (status, content_type, text). Raises _HttpError for 4xx/5xx shapes."""
split = urlsplit(target)
path = split.path
query = parse_qs(split.query)
if path == "/healthz":
if method != "GET":
raise _HttpError(_METHOD_NOT_ALLOWED, "GET only")
return 200, "application/json", json.dumps({"ok": True})
if path == "/memory/session-context":
if method != "GET":
raise _HttpError(_METHOD_NOT_ALLOWED, "GET only")
session_id = (query.get("session_id") or [None])[0]
params: dict = {}
if session_id:
params["session_id"] = session_id
# Per-call wake_depth override (minimal|standard|deep). standard/deep
# populate l0/l1/l2/rich_club -- the content worth injecting into a
# system prompt; core.py validates and ignores junk values.
wake_depth = (query.get("wake_depth") or [None])[0]
if wake_depth:
params["wake_depth"] = wake_depth
result = await self._dispatch("session_start_payload", params)
if (query.get("format") or ["json"])[0] == "text":
return 200, "text/plain; charset=utf-8", _payload_to_text(result)
return 200, "application/json", json.dumps(result, ensure_ascii=False)
if path == "/rpc":
if method != "POST":
raise _HttpError(_METHOD_NOT_ALLOWED, "POST only")
try:
envelope = json.loads(body or b"{}")
except json.JSONDecodeError as e:
raise _HttpError(_BAD_REQUEST, f"invalid json: {e}")
if not isinstance(envelope, dict) or not isinstance(envelope.get("method"), str):
raise _HttpError(_BAD_REQUEST, "expected {method, params}")
result = await self._dispatch(envelope["method"], envelope.get("params") or {})
return 200, "application/json", json.dumps(result, ensure_ascii=False)
raise _HttpError(_NOT_FOUND, f"no route for {path}")
async def _dispatch(self, method: str, params: dict) -> Any:
"""Run core.dispatch off-loop (R3) and map raises to _HttpError (R5)."""
try:
# CRITICAL R3: dispatch is sync + 50-500 ms; to_thread prevents
# head-of-line blocking across connections (same as socket_server).
return await asyncio.to_thread(dispatch, self.store, method, params)
except UnknownMethodError as e:
raise _HttpError(_NOT_FOUND, f"unknown method '{e.args[0]}'")
except KeyError as e:
raise _HttpError(_BAD_REQUEST, f"missing required param: {e.args[0]!r}")
except TypeError as e:
raise _HttpError(_BAD_REQUEST, str(e))
except Exception as e: # noqa: BLE001 -- HTTP must never crash daemon
raise _HttpError(_INTERNAL, str(e))
# -- connection handler -------------------------------------------------
async def handle(self, reader: asyncio.StreamReader, writer: asyncio.StreamWriter) -> None:
"""One coroutine per connection. Reads one request, responds, closes."""
self.active_connections += 1
t0 = time.monotonic()
method, target = "-", "-"
try:
self.last_activity_ts = t0
try:
parsed = await self._read_request(reader)
if parsed is None:
return
method, target, _headers, body = parsed
status, content_type, text = await self._route(method, target, body)
except _HttpError as e:
status, content_type, text = e.status, "application/json", json.dumps(
{"error": e.message}
)
except asyncio.IncompleteReadError:
status, content_type, text = _BAD_REQUEST, "application/json", json.dumps(
{"error": "truncated body"}
)
self._write_response(writer, status, content_type, text)
await writer.drain()
# Per-request access log -> stderr (journal). Matches the daemon's
# existing {"event":...} JSON-line style so it greps cleanly.
print(
json.dumps({
"event": "http_request",
"method": method,
"target": target,
"status": status,
"bytes": len(text.encode("utf-8")),
"ms": round((time.monotonic() - t0) * 1000),
}),
file=sys.stderr,
flush=True,
)
except (ConnectionResetError, BrokenPipeError, ConnectionAbortedError):
# Client hung up mid-write (fetch aborted, host killed). Expected;
# not a daemon fault -- mirrors socket_server.py's handling.
pass
finally:
self.active_connections -= 1
try:
writer.close()
await writer.wait_closed()
except Exception:
pass
@staticmethod
def _write_response(
writer: asyncio.StreamWriter, status: int, content_type: str, text: str
) -> None:
body = text.encode("utf-8")
reason = {
200: "OK", 400: "Bad Request", 404: "Not Found",
405: "Method Not Allowed", 500: "Internal Server Error",
}.get(status, "OK")
head = (
f"HTTP/1.0 {status} {reason}\r\n"
f"Content-Type: {content_type}\r\n"
f"Content-Length: {len(body)}\r\n"
"Connection: close\r\n"
"\r\n"
)
writer.write(head.encode("latin-1") + body)
# -- lifecycle ----------------------------------------------------------
async def serve(self) -> None:
"""Bind 127.0.0.1:port, write the port-file, serve until shutdown_event."""
server = await asyncio.start_server(self.handle, host=self.host, port=self.port)
self.bound_port = server.sockets[0].getsockname()[1]
# Discovery: publish the live port so consumers needn't guess. Best
# effort -- a failed write just means callers must be told the port.
try:
self.port_file.parent.mkdir(parents=True, exist_ok=True)
self.port_file.write_text(str(self.bound_port), encoding="utf-8")
os.chmod(self.port_file, 0o600)
except OSError:
pass
try:
async with server:
await self.shutdown_event.wait()
server.close()
await server.wait_closed()
finally:
try:
self.port_file.unlink()
except (FileNotFoundError, OSError):
pass
class _HttpError(Exception):
"""Internal control-flow exception carrying an HTTP status + message."""
def __init__(self, status: int, message: str) -> None:
super().__init__(message)
self.status = status
self.message = message

497
src/iai_mcp/qdrant_store.py
View file

@ -4,10 +4,10 @@ Replaces LanceDB with Qdrant for remote vector storage.
Collections (payload-partitioned, per Qdrant best practices):
- `records` : MemoryRecord rows (1024-dim cosine vectors)
All points carry `table: "records"` + `group_id` payload.
All points carry `table: "records"` + `group_id` payload.
- `metadata` : Payload-only (no vectors) containing edges, events,
budget_ledger, ratelimit_ledger.
Each point carries `table` + `group_id` payload.
budget_ledger, ratelimit_ledger.
Each point carries `table` + `group_id` payload.
Both collections use `is_tenant: true` payload indexes on `table` for
co-located storage and per-table HNSW graphs (payload_m=16, m=0).
@ -18,18 +18,6 @@ and data_json (events). AD = record UUID bytes.
The Qdrant client connects to a remote server via QDRANT_URL + QDRANT_API_KEY
environment variables (overridable via constructor).
Shim interface (_TableShim):
- ``add()`` routes by table name to dedicated upsert methods for all 5 tables
(EDGES_TABLE, EVENTS_TABLE, BUDGET_TABLE, RATELIMIT_TABLE, RECORDS_TABLE).
- ``to_pandas()`` constructs full DataFrames for edges, events, budget, and
ratelimit tables (records uses records_as_dataframe).
- ``_parse_where()`` translates LanceDB-style SQL clauses into Qdrant Filter:
``>`` / ``>=`` / ``<`` / ``<=`` emit Range conditions; ``=`` emits
MatchValue; multiple clauses joined with AND.
- ``events_query()`` scrolls all matching events and applies Python-side
filtering so events lacking ts_epoch (pre-migration rows) are never
silently dropped. Accepts both str and datetime for the since parameter.
"""
from __future__ import annotations
@ -644,7 +632,6 @@ class QdrantStore:
offset=offset,
scroll_filter=table_filter,
with_payload=True,
with_vectors=True,
)
all_points.extend(points)
if next_offset is None:
@ -652,64 +639,6 @@ class QdrantStore:
offset = next_offset
return [self._from_point(p) for p in all_points]
@staticmethod
def _parse_where(where: str | None) -> Filter | None:
"""Parse a LanceDB-style where clause into a Qdrant Filter.
Supported grammar:
- ``key = 'value'`` or ``key = "value"`` → MatchValue
- ``key > N`` or ``key >= N`` → Range(gt/gte)
- ``key < N`` or ``key <= N`` → Range(lt/lte)
- Multiple clauses joined by AND → Filter(must=[...])
Numeric values are auto-detected; string values must be
quoted. Returns None if the clause is empty or cannot be parsed.
"""
if not where:
return None
clauses = re.split(r'\s+AND\s+', where, flags=re.IGNORECASE)
conditions: list = []
for clause in clauses:
clause = clause.strip()
if not clause:
continue
m = re.match(
r"""([a-z_]+)\s*(>=|<=|>|<)\s*(-?\d+(?:\.\d+)?)\s*$""",
clause,
)
if m:
key, op, val_str = m.group(1), m.group(2), m.group(3)
val = float(val_str)
if op == ">":
conditions.append(FieldCondition(key=key, range=models.Range(gt=val)))
elif op == ">=":
conditions.append(FieldCondition(key=key, range=models.Range(gte=val)))
elif op == "<":
conditions.append(FieldCondition(key=key, range=models.Range(lt=val)))
elif op == "<=":
conditions.append(FieldCondition(key=key, range=models.Range(lte=val)))
continue
m = re.match(
r"""([a-z_]+)\s*=\s*['"]([^'"]+)['"]\s*$""",
clause,
)
if m:
key, val = m.group(1), m.group(2)
conditions.append(
FieldCondition(key=key, match=MatchValue(value=val))
)
continue
continue
if not conditions:
return None
return Filter(must=conditions)
def iter_records(
self,
*,
@ -720,10 +649,11 @@ class QdrantStore:
"""Streaming iterator over records (filtered by table=records)."""
offset = None
while True:
# Build filter: always include table=records, optionally add tier
conditions = [FieldCondition(key="table", match=MatchValue(value=RECORDS_TABLE))]
where_filter = self._parse_where(where)
if where_filter is not None:
conditions.extend(where_filter.must)
if where and where.startswith("tier = "):
tier = where.split("'")[1]
conditions.append(FieldCondition(key="tier", match=MatchValue(value=tier)))
qdrant_filter = Filter(must=conditions) if conditions else None
points, next_offset = self._client.scroll(
@ -731,7 +661,7 @@ class QdrantStore:
limit=batch_size,
offset=offset,
scroll_filter=qdrant_filter,
with_payload=columns if columns else True,
with_payload=True,
)
for point in points:
yield self._from_point(point)
@ -756,9 +686,9 @@ class QdrantStore:
offset = None
while True:
conditions = [FieldCondition(key="table", match=MatchValue(value=RECORDS_TABLE))]
where_filter = self._parse_where(where)
if where_filter is not None:
conditions.extend(where_filter.must)
if where and where.startswith("tier = "):
tier = where.split("'")[1]
conditions.append(FieldCondition(key="tier", match=MatchValue(value=tier)))
qdrant_filter = Filter(must=conditions) if conditions else None
points, next_offset = self._client.scroll(
@ -770,9 +700,9 @@ class QdrantStore:
with_vectors=False,
)
for point in points:
# Filter to requested columns
row = {k: v for k, v in point.payload.items() if k in columns}
if "id" in columns:
row["id"] = point.id
row["id"] = point.id
yield row
if next_offset is None:
@ -1269,7 +1199,6 @@ class QdrantStore:
"severity": severity,
"domain": domain,
"ts": ts.isoformat(),
"ts_epoch": ts.timestamp(),
"data_json": data_json,
"session_id": session_id,
"source_ids_json": source_ids_json,
@ -1277,14 +1206,9 @@ class QdrantStore:
)
self._client.upsert(collection_name=METADATA_TABLE, points=[point])
def events_query(self, kind: str | None = None, since: str | datetime | None = None,
def events_query(self, kind: str | None = None, since: datetime | None = None,
severity: str | None = None, limit: int = 100) -> list[dict]:
"""Query events from the metadata collection (table=events), newest first.
Scrolls all matching events and applies Python-side filtering so that
events lacking ``ts_epoch`` (pre-migration rows) are never silently
dropped.
"""
"""Query events from the metadata collection (table=events), newest first."""
# Always filter by table=events
conditions = [FieldCondition(key="table", match=MatchValue(value=EVENTS_TABLE))]
if kind is not None:
@ -1293,35 +1217,16 @@ class QdrantStore:
conditions.append(FieldCondition(key="severity", match=MatchValue(value=severity)))
event_filter = Filter(must=conditions) if conditions else None
points, _ = self._client.scroll(
collection_name=METADATA_TABLE,
limit=limit * 10, # fetch extra for post-filtering
with_payload=True,
with_vectors=False,
scroll_filter=event_filter,
)
# Resolve since to a datetime for Python-side comparison
since_dt: datetime | None = None
if since is not None:
if isinstance(since, str):
since_dt = datetime.fromisoformat(since)
else:
since_dt = since if since.tzinfo else since.replace(tzinfo=timezone.utc)
# Scroll ALL matching events (no limit on scroll — limit applies after sorting)
all_points: list = []
offset = None
while True:
points, next_offset = self._client.scroll(
collection_name=METADATA_TABLE,
limit=1000,
offset=offset,
with_payload=True,
with_vectors=False,
scroll_filter=event_filter,
)
all_points.extend(points)
if next_offset is None:
break
offset = next_offset
# Python-side filtering and sorting
out: list[dict] = []
for pt in all_points:
for pt in points:
p = pt.payload
ts_str = p.get("ts", "")
try:
@ -1329,17 +1234,16 @@ class QdrantStore:
except (ValueError, TypeError):
ts_dt = datetime.now(timezone.utc)
# Python-side since filter (works for all events, regardless of ts_epoch)
if since_dt is not None:
cmp = ts_dt if ts_dt.tzinfo else ts_dt.replace(tzinfo=timezone.utc)
if cmp < since_dt:
if since is not None:
since_cmp = since if since.tzinfo else since.replace(tzinfo=timezone.utc)
if ts_dt < since_cmp:
continue
raw_data = p.get("data_json") or "{}"
# data_json is AES-GCM encrypted; pass it through as a raw string
# so query_events() in events.py can decrypt it properly.
# Pre-02-08 unencrypted rows are plain JSON — query_events()
# handles both paths via is_encrypted() / decrypt_field().
try:
data = json.loads(raw_data)
except (TypeError, json.JSONDecodeError):
data = {}
try:
source_ids = json.loads(p.get("source_ids_json") or "[]")
except (TypeError, json.JSONDecodeError):
@ -1351,7 +1255,7 @@ class QdrantStore:
"severity": p.get("severity") or None,
"domain": p.get("domain") or None,
"ts": ts_dt,
"data": raw_data,
"data": data,
"session_id": p.get("session_id", "-"),
"source_ids": source_ids,
})
@ -1406,14 +1310,7 @@ class QdrantStore:
return pd.DataFrame(columns=["src", "dst", "edge_type", "weight", "updated_at"])
def records_as_dataframe(self) -> "pd.DataFrame":
"""Return all records from the records collection as a pandas DataFrame.
Column types match the LanceDB schema:
- timestamps are datetime objects (not str)
- structure_hv is bytes (not hex string)
- provenance_json is AES-256-GCM ciphertext (not plaintext JSON)
- tags_json is a JSON string
"""
"""Return all records from the records collection as a pandas DataFrame."""
try:
records = self.all_records()
if not records:
@ -1424,13 +1321,9 @@ class QdrantStore:
"stability", "difficulty", "last_reviewed",
"never_decay", "never_merge", "detail_level",
"s5_trust_score", "structure_hv",
"provenance_json", "created_at", "updated_at", "schema_version",
])
rows = []
for r in records:
# Re-encrypt provenance to match LanceDB's ciphertext column.
prov_plain = json.dumps(r.provenance or [])
prov_ct = self._encrypt_for_record(r.id, prov_plain)
rows.append({
"id": str(r.id),
"tier": r.tier,
@ -1439,21 +1332,17 @@ class QdrantStore:
"community_id": str(r.community_id) if r.community_id else None,
"centrality": r.centrality,
"pinned": r.pinned,
"tags_json": json.dumps(r.tags),
"tags_json": r.tags_json if hasattr(r, "tags_json") else "[]",
"language": r.language,
"aaak_index": r.aaak_index,
"stability": r.stability,
"difficulty": r.difficulty,
"last_reviewed": r.last_reviewed,
"last_reviewed": str(r.last_reviewed) if r.last_reviewed else None,
"never_decay": r.never_decay,
"never_merge": r.never_merge,
"detail_level": r.detail_level,
"s5_trust_score": r.s5_trust_score,
"structure_hv": bytes(r.structure_hv) if r.structure_hv else b"",
"provenance_json": prov_ct,
"created_at": r.created_at,
"updated_at": r.updated_at,
"schema_version": r.schema_version,
"structure_hv": r.structure_hv.hex() if r.structure_hv else "",
})
return pd.DataFrame(rows)
except Exception:
@ -1464,7 +1353,6 @@ class QdrantStore:
"stability", "difficulty", "last_reviewed",
"never_decay", "never_merge", "detail_level",
"s5_trust_score", "structure_hv",
"provenance_json", "created_at", "updated_at", "schema_version",
])
# ------------------------------------------------------------------ db shim
@ -1492,100 +1380,9 @@ class QdrantStore:
elif self._name == RECORDS_TABLE:
return self._store.records_as_dataframe()
elif self._name == EVENTS_TABLE:
return self._events_df()
elif self._name == BUDGET_TABLE:
return self._budget_df()
elif self._name == RATELIMIT_TABLE:
return self._ratelimit_df()
return pd.DataFrame()
return pd.DataFrame()
def _events_df(self) -> pd.DataFrame:
"""Return all events as a pandas DataFrame."""
try:
points = self._store._scroll_all(
METADATA_TABLE, table_filter=EVENTS_TABLE, batch_size=1000,
)
if not points:
return pd.DataFrame(columns=[
"id", "kind", "severity", "domain", "ts",
"data_json", "session_id", "source_ids_json",
])
rows = []
for pt in points:
p = pt.payload
ts_str = p.get("ts", "")
try:
ts_dt = datetime.fromisoformat(ts_str)
except (ValueError, TypeError):
ts_dt = datetime.fromisoformat("1970-01-01")
rows.append({
"id": p.get("id", str(pt.id)),
"kind": p.get("kind", ""),
"severity": p.get("severity") or "",
"domain": p.get("domain") or "",
"ts": ts_dt,
"data_json": p.get("data_json", ""),
"session_id": p.get("session_id", "-"),
"source_ids_json": p.get("source_ids_json", "[]"),
})
return pd.DataFrame(rows)
except Exception:
return pd.DataFrame(columns=[
"id", "kind", "severity", "domain", "ts",
"data_json", "session_id", "source_ids_json",
])
def _budget_df(self) -> pd.DataFrame:
"""Return all budget_ledger rows as a pandas DataFrame."""
try:
points = self._store._scroll_all(
METADATA_TABLE, table_filter=BUDGET_TABLE, batch_size=1000,
)
if not points:
return pd.DataFrame(columns=["date", "usd_spent", "kind", "ts"])
rows = []
for pt in points:
p = pt.payload
ts_str = p.get("ts", "")
try:
ts_dt = datetime.fromisoformat(ts_str)
except (ValueError, TypeError):
ts_dt = datetime.fromisoformat("1970-01-01")
rows.append({
"date": p.get("date", ""),
"usd_spent": float(p.get("usd_spent", 0.0)),
"kind": p.get("kind", ""),
"ts": ts_dt,
})
return pd.DataFrame(rows)
except Exception:
return pd.DataFrame(columns=["date", "usd_spent", "kind", "ts"])
def _ratelimit_df(self) -> pd.DataFrame:
"""Return all ratelimit_ledger rows as a pandas DataFrame."""
try:
points = self._store._scroll_all(
METADATA_TABLE, table_filter=RATELIMIT_TABLE, batch_size=1000,
)
if not points:
return pd.DataFrame(columns=["ts", "status_code", "endpoint"])
rows = []
for pt in points:
p = pt.payload
ts_str = p.get("ts", "")
try:
ts_dt = datetime.fromisoformat(ts_str)
except (ValueError, TypeError):
ts_dt = datetime.fromisoformat("1970-01-01")
rows.append({
"ts": ts_dt,
"status_code": int(p.get("status_code", 0)),
"endpoint": p.get("endpoint", ""),
})
return pd.DataFrame(rows)
except Exception:
return pd.DataFrame(columns=["ts", "status_code", "endpoint"])
def delete(self, where: str) -> None:
"""Delete rows from the table matching the where clause.
@ -1689,206 +1486,32 @@ class QdrantStore:
except Exception:
pass
def add(self, rows: list[dict]) -> None:
"""Insert rows into the table (LanceDB-compatible shim).
@staticmethod
def _parse_where(where: str | None) -> Filter | None:
"""Parse a LanceDB-style where clause into a Qdrant Filter.
LanceDB-compatible shim: ``table.add([{"col": "val", ...}])``.
Routes to the appropriate Qdrant collection based on table name.
Supported format: ``"key = 'value' AND key2 = 'value2'"``.
Each ``key = 'value'`` segment becomes a ``FieldCondition``
combined with ``must`` (AND semantics).
Returns None if the clause is empty or cannot be parsed.
"""
if not rows:
return
if self._name == EDGES_TABLE:
self._add_edges(rows)
elif self._name == EVENTS_TABLE:
self._add_events(rows)
elif self._name == BUDGET_TABLE:
self._add_budget(rows)
elif self._name == RATELIMIT_TABLE:
self._add_ratelimit(rows)
elif self._name == RECORDS_TABLE:
self._add_records(rows)
def _add_edges(self, rows: list[dict]) -> None:
points = []
for row in rows:
points.append(PointStruct(
id=str(uuid4()),
vector={},
payload={
"table": EDGES_TABLE,
"group_id": self._store._group_id,
"src": row.get("src", ""),
"dst": row.get("dst", ""),
"edge_type": row.get("edge_type", ""),
"weight": float(row.get("weight", 0.0)),
"updated_at": (
row["updated_at"].isoformat()
if isinstance(row.get("updated_at"), datetime)
else str(row.get("updated_at", ""))
),
},
))
if points:
self._store._client.upsert(
collection_name=METADATA_TABLE, points=points,
)
def _add_events(self, rows: list[dict]) -> None:
points = []
for row in rows:
ts_val = row.get("ts")
ts_str = (
ts_val.isoformat()
if isinstance(ts_val, datetime)
else str(ts_val) if ts_val else ""
)
# Compute ts_epoch for numeric range filtering (matches events_add)
ts_epoch = None
if isinstance(ts_val, datetime):
ts_epoch = ts_val.timestamp()
elif ts_str:
try:
ts_epoch = datetime.fromisoformat(ts_str).timestamp()
except (ValueError, TypeError):
pass
points.append(PointStruct(
id=str(row.get("id", uuid4())),
vector={},
payload={
"table": EVENTS_TABLE,
"group_id": self._store._group_id,
"id": str(row.get("id", "")),
"kind": row.get("kind", ""),
"severity": row.get("severity") or "",
"domain": row.get("domain") or "",
"ts": ts_str,
"ts_epoch": ts_epoch,
"data_json": row.get("data_json", ""),
"session_id": row.get("session_id", "-"),
"source_ids_json": row.get("source_ids_json", "[]"),
},
))
if points:
self._store._client.upsert(
collection_name=METADATA_TABLE, points=points,
)
def _add_budget(self, rows: list[dict]) -> None:
points = []
for row in rows:
ts_val = row.get("ts")
ts_str = (
ts_val.isoformat()
if isinstance(ts_val, datetime)
else str(ts_val) if ts_val else ""
)
points.append(PointStruct(
id=str(uuid4()),
vector={},
payload={
"table": BUDGET_TABLE,
"group_id": self._store._group_id,
"date": row.get("date", ""),
"usd_spent": float(row.get("usd_spent", 0.0)),
"kind": row.get("kind", ""),
"ts": ts_str,
},
))
if points:
self._store._client.upsert(
collection_name=METADATA_TABLE, points=points,
)
def _add_ratelimit(self, rows: list[dict]) -> None:
points = []
for row in rows:
ts_val = row.get("ts")
ts_str = (
ts_val.isoformat()
if isinstance(ts_val, datetime)
else str(ts_val) if ts_val else ""
)
points.append(PointStruct(
id=str(uuid4()),
vector={},
payload={
"table": RATELIMIT_TABLE,
"group_id": self._store._group_id,
"ts": ts_str,
"status_code": int(row.get("status_code", 0)),
"endpoint": row.get("endpoint", ""),
},
))
if points:
self._store._client.upsert(
collection_name=METADATA_TABLE, points=points,
)
def _add_records(self, rows: list[dict]) -> None:
"""Insert record rows — builds PointStruct directly from row dicts.
The row dicts come from _to_row() which already contains
encrypted ciphertext for literal_surface / provenance_json /
profile_modulation_gain_json. Building a MemoryRecord and
passing it through _to_point would double-encrypt those fields,
so we construct the PointStruct payload directly here.
"""
points = []
for row in rows:
def _ts(val):
if val is None:
return None
if isinstance(val, datetime):
return val.isoformat()
return str(val)
# Decode structure_hv from bytes (LanceDB pa.binary → bytes)
structure_raw = row.get("structure_hv")
if isinstance(structure_raw, (bytes, bytearray)):
structure_b64 = base64.b64encode(structure_raw).decode("ascii")
elif isinstance(structure_raw, str):
# Already base64 or hex — keep as-is
structure_b64 = structure_raw
else:
structure_b64 = ""
points.append(PointStruct(
id=str(row.get("id", uuid4())),
vector=list(row.get("embedding", [0.0] * self._store._embed_dim)),
payload={
"table": RECORDS_TABLE,
"group_id": self._store._group_id,
"tier": row.get("tier", "episodic"),
"literal_surface": row.get("literal_surface", ""),
"aaak_index": row.get("aaak_index", ""),
"structure_hv": structure_b64,
"community_id": row.get("community_id", ""),
"centrality": float(row.get("centrality", 0.0)),
"detail_level": int(row.get("detail_level", 1)),
"pinned": bool(row.get("pinned", False)),
"stability": float(row.get("stability", 0.0)),
"difficulty": float(row.get("difficulty", 0.0)),
"last_reviewed": _ts(row.get("last_reviewed")),
"never_decay": bool(row.get("never_decay", False)),
"never_merge": bool(row.get("never_merge", False)),
"provenance_json": row.get("provenance_json", "[]"),
"created_at": _ts(row.get("created_at")),
"updated_at": _ts(row.get("updated_at")),
"tags_json": row.get("tags_json", "[]"),
"language": str(row.get("language", "en")),
"s5_trust_score": float(row.get("s5_trust_score", 0.5)),
"profile_modulation_gain_json": row.get("profile_modulation_gain_json", "{}"),
"schema_version": int(row.get("schema_version", 1)),
},
))
if points:
self._store._client.upsert(
collection_name=RECORDS_TABLE, points=points,
)
def _parse_where(self, where: str | None) -> Filter | None:
"""Delegate to QdrantStore._parse_where (single source of truth)."""
return QdrantStore._parse_where(where)
if not where:
return None
# Find all key = 'value' or key = "value" patterns.
pairs = re.findall(
r"""([a-z_]+)\s*=\s*['"]([^'"]+)['"]""",
where,
)
if not pairs:
return None
conditions = [
FieldCondition(key=key, match=MatchValue(value=val))
for key, val in pairs
]
if len(conditions) == 1:
return Filter(must=conditions)
return Filter(must=conditions)
@property
def db(self) -> "QdrantStore._DbShim":

6
src/iai_mcp/retrieve.py
View file

@ -278,12 +278,6 @@ def contradict(
store.insert(new_rec)
store.add_contradicts_edge(original_id, new_rec.id)
# Wire temporal_next: contradict inserts land in same session, link them.
try:
link_temporal_next(store, new_rec, session_id="-")
except Exception:
pass # diagnostic only; never block the write path.
# monotropic proactive check fires only in high-focus
# domains. Hints aren't surfaced via contradict() (its signature is fixed
# to ReconsolidationReceipt), but events land in the events table so the

14
src/iai_mcp/sleep.py
View file

@ -296,22 +296,16 @@ def run_light_consolidation(
def _build_hebbian_clusters(store: MemoryStore) -> list[list[UUID]]:
"""Find connected components in the hebbian+temporal_next edge graph
with size >= CLUSTER_MIN_SIZE.
Hebbian edges capture semantic similarity (dedup/reinforce);
temporal_next edges capture sequential proximity (same-session inserts
within 5 minutes). Both are valid signals for CLS clustering.
"""
"""Find connected components in the hebbian edge graph with size >= CLUSTER_MIN_SIZE."""
edges_df = store.db.open_table(EDGES_TABLE).to_pandas()
if edges_df.empty:
return []
relevant = edges_df[edges_df["edge_type"].isin(("hebbian", "temporal_next"))]
if relevant.empty:
hebbian = edges_df[edges_df["edge_type"] == "hebbian"]
if hebbian.empty:
return []
adj: dict[UUID, set[UUID]] = {}
for _, row in relevant.iterrows():
for _, row in hebbian.iterrows():
src = UUID(row["src"])
dst = UUID(row["dst"])
adj.setdefault(src, set()).add(dst)

248
tests/test_http_server.py
View file

@ -1,248 +0,0 @@
"""End-to-end tests for the Tier-1 localhost HTTP adapter (http_server.py).
Mirrors tests/test_daemon_dispatcher.py: boot the REAL HttpServer on an
ephemeral 127.0.0.1 port and drive it with raw HTTP/1.0 over a real TCP
socket. core.dispatch is monkeypatched to a fake so these tests exercise the
*transport adapter* (routing, error->status mapping, port-file discovery,
graceful shutdown) without needing a real MemoryStore / LanceDB.
"""
from __future__ import annotations
import asyncio
import json
import pytest
from iai_mcp.core import UnknownMethodError
def _fake_dispatch(store, method, params):
"""Stand-in for core.dispatch: deterministic, no store needed."""
if method == "session_start_payload":
return {
"l0": "pinned identity",
"l1": "recent summary",
"l2": ["episode one", "episode two"],
"rich_club": "hub concepts",
"total_cached_tokens": 0,
"total_dynamic_tokens": 1000,
# echo the override so the transport-threading test can assert it.
"wake_depth": params.get("wake_depth", "minimal"),
}
if method == "minimal_payload":
# wake_depth=minimal: content layers empty, only opaque handles set.
return {
"l0": "", "l1": "", "l2": [], "rich_club": "",
"identity_pointer": "",
"brain_handle": "<sess:abc pend:2>",
"topic_cluster_hint": "<topic:d3256b25>",
"compact_handle": "<iai:970d1b7629da7946>",
"wake_depth": "minimal",
}
if method == "needs_param":
# Simulate core.py's params["cue"] KeyError path.
return {"cue": params["cue"]}
if method == "boom":
raise RuntimeError("kaboom")
raise UnknownMethodError(method)
async def _http_request(port, method, path, body=None, *, timeout=5.0):
"""Send one HTTP/1.0 request, read the full response to EOF, parse it."""
reader, writer = await asyncio.wait_for(
asyncio.open_connection("127.0.0.1", port), timeout=timeout
)
try:
body_bytes = body.encode("utf-8") if isinstance(body, str) else (body or b"")
head = f"{method} {path} HTTP/1.0\r\n"
if body_bytes:
head += f"Content-Length: {len(body_bytes)}\r\n"
head += "Connection: close\r\n\r\n"
writer.write(head.encode("latin-1") + body_bytes)
await writer.drain()
raw = await asyncio.wait_for(reader.read(-1), timeout=timeout)
finally:
writer.close()
try:
await writer.wait_closed()
except Exception:
pass
head_bytes, _, payload = raw.partition(b"\r\n\r\n")
lines = head_bytes.split(b"\r\n")
status = int(lines[0].decode("latin-1").split(" ")[1])
headers = {}
for line in lines[1:]:
name, _, val = line.decode("latin-1").partition(":")
headers[name.strip().lower()] = val.strip()
return status, headers, payload.decode("utf-8")
async def _with_http_server(coro_fn, *, port_file, monkeypatch):
"""Boot the real HttpServer with _fake_dispatch on an ephemeral port."""
from iai_mcp import http_server as hs
monkeypatch.setattr(hs, "dispatch", _fake_dispatch)
server = hs.HttpServer(store=object(), host="127.0.0.1", port=0, port_file=port_file)
task = asyncio.create_task(server.serve())
for _ in range(250): # wait for bind (bound_port set inside serve())
if server.bound_port is not None:
break
await asyncio.sleep(0.01)
if server.bound_port is None:
server.shutdown_event.set()
await asyncio.wait_for(task, timeout=5)
raise AssertionError("server never bound")
try:
return await coro_fn(server)
finally:
server.shutdown_event.set()
try:
await asyncio.wait_for(task, timeout=5)
except Exception:
pass
def test_healthz_returns_ok(tmp_path, monkeypatch):
async def _runner(server):
return await _http_request(server.bound_port, "GET", "/healthz")
status, headers, body = asyncio.run(
_with_http_server(_runner, port_file=tmp_path / ".http.port", monkeypatch=monkeypatch)
)
assert status == 200
assert headers["content-type"] == "application/json"
assert json.loads(body) == {"ok": True}
def test_session_context_json(tmp_path, monkeypatch):
async def _runner(server):
return await _http_request(
server.bound_port, "GET", "/memory/session-context?session_id=abc"
)
status, headers, body = asyncio.run(
_with_http_server(_runner, port_file=tmp_path / ".http.port", monkeypatch=monkeypatch)
)
assert status == 200
assert headers["content-type"] == "application/json"
payload = json.loads(body)
assert payload["l0"] == "pinned identity"
assert payload["l2"] == ["episode one", "episode two"]
def test_session_context_threads_wake_depth(tmp_path, monkeypatch):
async def _runner(server):
return await _http_request(
server.bound_port,
"GET",
"/memory/session-context?session_id=abc&wake_depth=standard",
)
status, _headers, body = asyncio.run(
_with_http_server(_runner, port_file=tmp_path / ".http.port", monkeypatch=monkeypatch)
)
assert status == 200
assert json.loads(body)["wake_depth"] == "standard"
def test_session_context_text_format(tmp_path, monkeypatch):
async def _runner(server):
return await _http_request(
server.bound_port, "GET", "/memory/session-context?format=text"
)
status, headers, body = asyncio.run(
_with_http_server(_runner, port_file=tmp_path / ".http.port", monkeypatch=monkeypatch)
)
assert status == 200
assert headers["content-type"].startswith("text/plain")
# Flattened block keeps the human layers in order, drops token ints.
assert body == (
"## L0 Identity\npinned identity\n\n## L1 Recent Summary\nrecent summary\n\n"
"## Rich Club\nhub concepts\n\n## L2 Episode\nepisode one\n\n## L2 Episode\nepisode two"
)
def test_payload_to_text_renders_handles_at_minimal():
"""At minimal wake_depth, _payload_to_text surfaces the compact handles
(content layers empty) so the injected block is never silently empty."""
from iai_mcp.http_server import _payload_to_text
text = _payload_to_text(_fake_dispatch(None, "minimal_payload", {}))
assert text == "## Handles\n<iai:970d1b7629da7946> <sess:abc pend:2> <topic:d3256b25>"
def test_rpc_post_passthrough(tmp_path, monkeypatch):
async def _runner(server):
return await _http_request(
server.bound_port,
"POST",
"/rpc",
json.dumps({"method": "session_start_payload", "params": {"session_id": "z"}}),
)
status, _headers, body = asyncio.run(
_with_http_server(_runner, port_file=tmp_path / ".http.port", monkeypatch=monkeypatch)
)
assert status == 200
assert json.loads(body)["rich_club"] == "hub concepts"
def test_unknown_route_404(tmp_path, monkeypatch):
async def _runner(server):
return await _http_request(server.bound_port, "GET", "/nope")
status, _headers, body = asyncio.run(
_with_http_server(_runner, port_file=tmp_path / ".http.port", monkeypatch=monkeypatch)
)
assert status == 404
assert "no route" in json.loads(body)["error"]
def test_unknown_method_maps_to_404(tmp_path, monkeypatch):
async def _runner(server):
return await _http_request(
server.bound_port, "POST", "/rpc", json.dumps({"method": "ghost"})
)
status, _headers, body = asyncio.run(
_with_http_server(_runner, port_file=tmp_path / ".http.port", monkeypatch=monkeypatch)
)
assert status == 404
assert "unknown method 'ghost'" in json.loads(body)["error"]
def test_internal_error_maps_to_500(tmp_path, monkeypatch):
async def _runner(server):
return await _http_request(
server.bound_port, "POST", "/rpc", json.dumps({"method": "boom"})
)
status, _headers, body = asyncio.run(
_with_http_server(_runner, port_file=tmp_path / ".http.port", monkeypatch=monkeypatch)
)
assert status == 500
assert "kaboom" in json.loads(body)["error"]
def test_wrong_verb_405(tmp_path, monkeypatch):
async def _runner(server):
return await _http_request(server.bound_port, "POST", "/healthz")
status, _headers, _body = asyncio.run(
_with_http_server(_runner, port_file=tmp_path / ".http.port", monkeypatch=monkeypatch)
)
assert status == 405
def test_port_file_written_and_cleaned(tmp_path, monkeypatch):
port_file = tmp_path / ".http.port"
async def _runner(server):
# While serving, the port-file holds the live bound port.
assert port_file.read_text() == str(server.bound_port)
return server.bound_port
asyncio.run(_with_http_server(_runner, port_file=port_file, monkeypatch=monkeypatch))
# After graceful shutdown, the discovery file is removed.
assert not port_file.exists()

58
tests/test_session_start_wake_depth.py
View file

@ -1,58 +0,0 @@
"""Unit tests for the per-call wake_depth override in core.dispatch.
The HTTP adapter (and any caller) can request standard/deep session-start
content without flipping the global profile knob. These tests verify the
override is threaded into assemble_session_start's profile_state, that invalid
values fall back to the per-process state, and that the global _profile_state is
never mutated. assemble_session_start + build_runtime_graph are monkeypatched
so no real store/embedder is needed.
"""
from __future__ import annotations
import iai_mcp.core as core
import iai_mcp.retrieve as retrieve
import iai_mcp.session as session
class _FakeStore:
def count_rows(self, table): # non-empty -> takes the assemble branch
return 5
def _patch(monkeypatch):
captured = {}
def fake_assemble(store, assignment, rc, *, session_id, profile_state):
captured["profile_state"] = profile_state
captured["session_id"] = session_id
return session.SessionStartPayload(l0="content")
monkeypatch.setattr(session, "assemble_session_start", fake_assemble)
monkeypatch.setattr(retrieve, "build_runtime_graph", lambda store: (None, {}, None))
monkeypatch.setattr(core, "_profile_state", {"wake_depth": "minimal", "literal_preservation": 0.5})
return captured
def test_wake_depth_override_threaded(monkeypatch):
captured = _patch(monkeypatch)
core.dispatch(_FakeStore(), "session_start_payload", {"session_id": "s", "wake_depth": "standard"})
ps = captured["profile_state"]
assert ps["wake_depth"] == "standard"
# other profile knobs are preserved in the override copy.
assert ps["literal_preservation"] == 0.5
# the module global must NOT be mutated by the override.
assert core._profile_state["wake_depth"] == "minimal"
def test_invalid_wake_depth_falls_back_to_profile(monkeypatch):
captured = _patch(monkeypatch)
core.dispatch(_FakeStore(), "session_start_payload", {"session_id": "s", "wake_depth": "ultra"})
# junk value ignored -> uses the per-process profile state (identity, minimal).
assert captured["profile_state"] is core._profile_state
assert captured["profile_state"]["wake_depth"] == "minimal"
def test_absent_wake_depth_uses_profile(monkeypatch):
captured = _patch(monkeypatch)
core.dispatch(_FakeStore(), "session_start_payload", {"session_id": "s"})
assert captured["profile_state"] is core._profile_state