wire the broker to the compositor, containers and pointer

The four slices existed and nothing called them. Now the broker builds a
spatial container per session, reads back what the compositor actually did,
and aims the cursor into one container under a broker-wide lease.

Record schema 3 (registry.py): `container` {output, workspace, mode,
created_by, state}, `pointer` {bookmark, last_released_at, released_by} and
the compositor's `signature`. A v2 record is migrated on load, not refused —
its application is still running. Every field that could not be verified is
null; `container: null` means "no container", never "probably something".

Containers (new container_host.py, manager.py, launcher.py): `start`
allocates a slot from `WorkspaceAllocator`, runs `container_plan()`'s argv,
launches the app with the container stamp in its environment, then reads the
workspace, output and window back out of `clients -j`. Teardown removes the
output, releases the slot and reports survivors rather than claiming success
— and runs for an unprompted exit and for `gc` too, so no container is ever
left unattended. Broker start-up reconciles: adopt `hh-*` outputs a live
session claims, remove the rest, `contains_output()` being the only filter.
No IPC means `container: degraded` and lifecycle verbs unchanged.

The plans' argv are dispatched over the compositor's request socket rather
than by executing `hyprctl`, which is what hyprctl does with the same words;
the reply is checked, so a refused command is not recorded as a container.

Pointer (new driver.py, manager.py): a client for the driver's private
socket, and `pointer.acquire|move|button|click|release|state` on top of
`PointerLease`. The broker owns the lease, the driver executes. The bookmark
is marked on every move and before release and stored in the record, so it
survives a restart. On expiry `pressed()` is read before `expire()` and
`release_all` goes to the driver, because the lease cannot lift a button
itself; a sweep runs unprompted so the TTL is a promise.

containers.py to v2: `quarantine_plan()` -> `quarantine_argv()`, with the
before-any-per-session-rule ordering stated where the caller could get it
wrong. protocol.py: 9 NO_LEASE, 10 COMPOSITOR, the verb table, and the
`request_id` cache moved out of the server so the CLI agrees with it.

Tests: real AF_UNIX everywhere — a stateful fake compositor and the shipped
pointer driver behind a recording sink. No mocked sockets, no compositor, no
VM. 551 -> 693.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017mUq7QvQ9EZb2Wi35xj4cs
This commit is contained in:
HermenixandClaude Opus 5 committed 2026-10-02 19:47:33 +00:00
1 parent 20b2aee109
commit 129430bffd
20 files changed
+4805 -89

No files matched your search

+114
View File
@@ -9,6 +9,7 @@ rest of ``hhstreamd``.
from __future__ import annotations
import collections
import enum
import json
from typing import Any
@@ -48,6 +49,14 @@ class ExitCode(enum.IntEnum):
POLICY = 6
TIMEOUT = 7
BUSY = 8
#: The verb needs the pointer lease and this session does not hold it.
#: Distinct from BUSY, which is "someone else holds it": an agent that
#: sees 9 must acquire, one that sees 8 must wait.
NO_LEASE = 9
#: An IPC or output operation failed. **The session is not lost** — it
#: keeps running in degraded mode — so this must never be confused with a
#: session failure, which is what a shared code would do.
COMPOSITOR = 10
#: Machine-readable error names, grouped by the exit code they map to. Every
@@ -80,6 +89,9 @@ EXIT_CODE_FOR_ERROR: dict[str, ExitCode] = {
"not_allowlisted": ExitCode.POLICY,
"argument_refused": ExitCode.POLICY,
"session_cap_reached": ExitCode.POLICY,
# The workspace band is full. A policy refusal and not a compositor
# failure: the compositor is fine, the broker is at its budget.
"container_budget": ExitCode.POLICY,
"refused": ExitCode.POLICY,
# 7 — timed out
"timeout": ExitCode.TIMEOUT,
@@ -87,6 +99,15 @@ EXIT_CODE_FOR_ERROR: dict[str, ExitCode] = {
# 8 — busy
"busy": ExitCode.BUSY,
"already_managed": ExitCode.BUSY,
"pointer_busy": ExitCode.BUSY,
# 9 — the lease is required and not held
"no_lease": ExitCode.NO_LEASE,
# 10 — the compositor side failed; the session itself is fine
"compositor": ExitCode.COMPOSITOR,
"compositor_unavailable": ExitCode.COMPOSITOR,
"no_container": ExitCode.COMPOSITOR,
"driver_unavailable": ExitCode.COMPOSITOR,
"driver_refused": ExitCode.COMPOSITOR,
}
@@ -103,6 +124,99 @@ def exit_code_for(error: str | None) -> ExitCode:
return EXIT_CODE_FOR_ERROR.get(error, ExitCode.BAD_STATE)
# --------------------------------------------------------------------------
# Verbs and `request_id` idempotency (PLAN §8)
# --------------------------------------------------------------------------
#: Every verb the daemon answers. The container and pointer verbs are dotted
#: and the original eight are not, because the original eight are frozen
#: (CONTRACT-sessions.md §4: "existing verbs are unchanged") and renaming them
#: would break every agent for a cosmetic gain.
VERBS = (
"ping",
"start",
"list",
"status",
"touch",
"kill",
"gc",
"adopt",
"logs",
"container.list",
"container.reconcile",
"pointer.acquire",
"pointer.move",
"pointer.button",
"pointer.click",
"pointer.release",
"pointer.state",
)
#: Verbs whose reply is remembered by ``request_id``, so a caller that died
#: before reading the reply can retry without the mutation happening twice.
#:
#: The read verbs are absent deliberately: they are idempotent already, so a
#: retry costs nothing, and a cached ``list`` would be a stale one. Every
#: ``pointer.*`` verb that moves something is here, because a retried click is
#: a second click — the one failure an agent cannot undo.
MUTATING_VERBS = frozenset(
{
"start",
"touch",
"kill",
"adopt",
"gc",
"container.reconcile",
"pointer.acquire",
"pointer.move",
"pointer.button",
"pointer.click",
"pointer.release",
}
)
#: How many completed replies are kept. Bounded because a client chooses the
#: ids: an unbounded cache is memory a peer controls.
DEDUP_ENTRIES = 64
class IdempotencyCache:
"""The last N mutating replies, by ``request_id``.
A retry returns *the recorded answer*, including a recorded failure: an
agent that retries a ``start`` which was refused by policy must see the
refusal again, not a second attempt that might succeed and leave it with
an application it has no id for. Replies are returned with
``deduplicated: true`` so the caller can tell a replay from a fresh answer
rather than having to infer it.
"""
def __init__(self, capacity: int = DEDUP_ENTRIES) -> None:
self.capacity = capacity
self._entries: collections.OrderedDict[str, dict[str, Any]] = collections.OrderedDict()
def lookup(self, op: str, request_id: Any) -> dict[str, Any] | None:
"""The recorded reply for this id, or ``None`` to go ahead and do it."""
if not isinstance(request_id, str) or op not in MUTATING_VERBS:
return None
recorded = self._entries.get(request_id)
if recorded is None:
return None
return {**recorded, "deduplicated": True}
def remember(self, op: str, request_id: Any, response: dict[str, Any]) -> None:
if not isinstance(request_id, str) or op not in MUTATING_VERBS:
return
self._entries[request_id] = dict(response)
self._entries.move_to_end(request_id)
while len(self._entries) > self.capacity:
self._entries.popitem(last=False)
def __len__(self) -> int:
return len(self._entries)
# --------------------------------------------------------------------------
# Message construction
# --------------------------------------------------------------------------