boxmaker/docs/plans/M5a/README.md

M5a implementation plan: local memory

For the implementing model: do not work from this file. The owner gives you one task file at a time (01-… to 20-…). This file is the index for the owner and the reviewer.

Goal: a new role, memoryd, owns the agent's memory: identity (soul.md, memory.md, user.md, in every new session's prompt) and experience (notes.md and the session logs, found with recall). The agent writes memory with a memory tool without asking; a session that has read untrusted content can only take notes. brokerd decides both tools under built-in grants and hands them to memoryd; loopd takes the memory block from memoryd for each new session.

Architecture: brokerd's two largest files are split first (01). proto gains the tools' arguments and eight message kinds (02). memoryd is built bottom up: its crate, configuration and sockets (03), entries and the check (04), the files and the block (05), notes and the edits (06), the journal, undo and review (07), the index (08), recall (09), and serving (10). brokerd gains the two tools and their built-in grants (11), then sends decided memory calls to memoryd with the session's state, keeping the stricter label (12). loopd takes the block and declares the two core tools (13); bxctl gains reindex and memory (14). Last, the starter soul.md and the device checks with memoryd running (15).

Spec: docs/specs/2026-09-24-m5a-local-memory.md (approved; P17 to P21 applied to the brief). Runbook: task 03 adds memoryd-start-failed, task 12 memoryd-unreachable, task 13 replaces core-memory-unreadable with memory-unavailable and adds memory-index-damaged.

New dependency: rusqlite 0.40.2 with bundled (task 03; deny.toml gains Zlib for foldhash). Nothing else.

Global constraints

• Everything in AGENTS.md, including "Lessons from earlier reviews".
• Branch m5a. One task, one fresh OpenCode session, one commit. Run cargo fmt --all before the gate. Review happens once, after task 15.
• New files come as skeletons: types, constants and the glue written, todo!() bodies with the code to write in the comment above. Existing files are changed by the exact edits in the task, never replaced whole.
• Some tasks hand over new versions of given test files, because the change is in what the test must see. Copy them over the old ones as the task says; they are still not to be edited.
• Only task 15 needs straylight (make verify-device). Everything else runs offline against temporary directories and fakes.

Tasks

Every task had a reference implementation, built in this order from master with the gate at every step and replayed on its own (tip T26), its given tests run against it, and mutations of it caught by them (tip T28); then the reference was moved off this machine (tip T18). Every skeleton was checked to compile at its task's end state and to fail its tests, and every comment above a todo!() was read beside its signature (tip T29). The three largest functions were split before the skeletons were cut (tip T25).

#	File	Delivers	Tests	Check
01	01-brokerd-split.md	policy/ and ledger/ as modules; no behaviour change	the existing ones	reference
02	02-proto-memory.md	the tools' arguments, eight message kinds	memory_args.rs, memory_wire.rs, 8 fixtures	reference; 5 mutations caught
03	03-memoryd-crate.md	the crate, configuration, listen, rusqlite	config.rs, socket.rs	reference; cargo deny
04	04-memoryd-entries-check.md	entries; the check given complete	entries.rs, check.rs	reference; 8 mutations caught
05	05-memoryd-files.md	the files, the block	files.rs	reference; 6 mutations caught
06	06-memoryd-edits.md	notes, the five actions, routing by trust	edit.rs, notes.rs	reference; 11 mutations caught
07	07-memoryd-journal.md	the journal, undo, review	journal.rs	reference; 5 mutations caught
08	08-memoryd-index.md	the FTS5 index, labels, catch-up, rebuild	index.rs, support/logs.rs	reference; 8 mutations caught
09	09-memoryd-recall.md	recall, source tags, the recall log	recall.rs	reference; 8 mutations caught
10	10-memoryd-serve.md	the three sockets; memoryd serve	serve.rs	reference; 10 mutations caught
11	11-brokerd-memory-tools.md	the two tools, built-in grants	memory_tools.rs; grants.rs updated	reference; 6 mutations caught
12	12-brokerd-memory-port.md	calls to memoryd, stricter labels	memory_port.rs; rig.rs updated	reference; 8 mutations caught
13	13-loopd-memory.md	the block, the core tools, session_updated	memory.rs; baseline.rs, pointers.rs, channel.rs updated	reference; 6 mutations caught
14	14-bxctl-memory.md	bxctl reindex, bxctl memory	cli_memory.rs	reference; 5 mutations caught
15	15-soul-and-device.md	the starter soul; device checks with memoryd	device.rs updated, support/memoryd.rs	reference; make verify-device on straylight
16	16-macos-gate.md	the gate passes on macOS: tests use /tmp, gatewayd finds its uid without /proc	shell.rs updated	reference; gate on titan and talos; 1 mutation caught (7 tests)
17	17-memoryd-no-slicing.md	the day and the month with get, not a slice (review nit)	the existing ones	reference
18	18-repeated-code.md	repeated error lines written once (review nit)	the existing ones	reference; 4 mutations caught
19	19-no-forged-headers.md	a note cannot forge a notes.md header (review finding)	forgery.rs; memory_args.rs, notes.rs updated	reference; 4 mutations caught
20	20-review-ids.md	a staged write never reuses an id (review finding)	review_ids.rs	reference; 2 mutations caught
21	21-session-follows-result.md	a session's state follows the result's own label (device finding)	ledger_label.rs	reference; 5 mutations caught
22	22-plain-note-answers.md	the answers to a note say where it went (device finding)	edit.rs, serve.rs updated	reference; 4 mutations caught

At the end of task 15: 936 tests (854 before task 01); tasks 16 to 18 add no test and keep 936; at the end of task 19, 939; at the end of task 20, 941; at the end of task 21, 944, and task 22 adds none. Each task's count was checked by running the gate at its end state on its own.

Changes during the run

• 2026-09-24, task 04, first stop: tasks 01 to 03 committed, each on its first session. Task 04's session read the given test and the skeleton, which held chat-template tags literally, because the check refuses them. Ornith took them for its own template. Its reasoning ended mid-thought, it lost track of the file, and it made no change. The tags were then built from pieces in check.rs and the four given tests that held them, and the spec describes them in words (e40906f).
• 2026-09-24, task 04, second stop: this time Ornith copied the files and saw the tests fail, then checked its planned code against the markup test cases in its reasoning, spelling out the tags as it went. The tag it wrote ended its reasoning; what followed became its answer, and the turn ended with no tool call. Building the tags from pieces keeps them out of files, not out of the model's thoughts. So check.rs is now given complete (the reference's, tags from pieces), and task 04 fills only entries.rs and says not to work through the markup cases. The three later tests that used a tag only as something the check refuses (edit.rs, files.rs, serve.rs) now use an instruction phrase or an invisible character. They pass against the reference and are rustfmt-clean. The attempt is saved in .state/runs/M5a/04-attempt-2/. Resume from task 04. The measurements behind this are in docs/research/2026-09-24-chat-template-tokens.md, added after the run.
• 2026-09-24, after the review: tasks 17 and 18 clean up the review's nits. Task 19 closes a hole the review found in the spec and the reference alike: a note's text could hold a line shaped like a notes.md header, so an untrusted session could plant a moved from memory.md entry that recall showed as identity. The spec was changed (4ae827a, section 3). Since task 19 hands over new versions of two given tests, task 02's memory_args.rs is now memory_args.rs-task02 and task 06's notes.rs is now notes.rs-task06. Task 20 fixes review ids: counting only the waiting writes gave a write staged after an accepted one the accepted one's id, and the write was lost.
• 2026-09-24, later: the work moved to titan (macOS) while talos's disk is replaced, and there the gate failed for reasons older than M5a: socket paths under macOS's long $TMPDIR, gatewayd reading /proc/self, and /tmp being /private/tmp. Task 16 makes it pass on both; the review's tasks moved to 17 to 20. Resume from task 16.
• 2026-09-25, the device check on straylight: a clean session that recalled a memory-tainted note stayed trusted, because the ledger raised the state by the grant's label alone, and it wrote the note into memory.md. The spec said "as today" and the reference did the same, so task 21 fixes both (0e5b1d2, section 8). And the model twice told the owner "Added to memory.md" for a write that had become a note; task 22 makes the answers say where it went (section 3). Task 22 hands over new versions of two given tests: task 06's edit.rs is now edit.rs-task06 and task 10's serve.rs is now serve.rs-task10. Resume from task 21.

Running it

git switch m5a
OPENCODE_MODEL=straylight/ornith-1.5-35b-a3b tools/run-plan.sh docs/plans/M5a
# the review's follow-ups, on titan, which serves Ornith itself (127.0.0.1:8081):
OPENCODE_MODEL=titan/ornith-1.5-35b-a3b tools/run-plan.sh docs/plans/M5a 16
# the device check's follow-ups:
OPENCODE_MODEL=titan/ornith-1.5-35b-a3b tools/run-plan.sh docs/plans/M5a 21

The model the implementer runs on is separate from the one the harness serves: make verify-device and the device checks still use straylight's llama-server.

Keep the OpenCode TUI closed while it runs.