boxmaker/docs/plans/M5d1

M5d-1 implementation plan: epochs and compaction

For the implementer: each task file is one closed task for one fresh session. Read AGENTS.md, then the task file, and do exactly that task. The driver is tools/run-plan.sh.

Goal: a live session can start a new epoch. loopd writes N.jsonl / N.baseline.json for N > 0, turn numbers and untrusted() span every epoch, compaction runs after a turn whose last Usage is at or above a configured fraction of expect.n_ctx, the summary call uses the background slot, and memoryd's Honcho walker reads every epoch file of a session.

Architecture: loopd gains a [compact] table and a compact module (when to compact, the summary call, writing the next epoch). Session::open loads the highest complete epoch and the carried summary. messages for epoch n > 0 inserts that carry as a ChatMessage::User that is never a LogRecord::User. Honcho concatenates epoch logs in order so Item::Turn numbers stay monotonic. Heartbeat, cron, ToolSummary, and tiering are M5d-2.

Tech stack: Rust 1.98.1, edition 2024; no new crate.

Spec: docs/specs/2026-10-03-m5d-scheduler.md, sections 3, 5 ([compact] without using summarise), 7 (the M5d-1 tests), and 8 (session-compact-failed).

The owner's decisions (2026-10-03)

1. Compaction runs after a turn, when that turn's cache_n + prompt_n is at or above a configured fraction of expect.n_ctx (default 0.85). A quiet session is not a reason to compact.
2. If a live request is exceed_context_size_error, compact and retry that user message once in the new epoch.
3. The summary call is on the background slot, with no tools, thinking on. A context-full summary falls back to the epoch's User texts, newest last, cut to summary_chars.
4. The carried summary is not a LogRecord::User, so it is not a Honcho turn.

What the code already has, and what it lacks

• LogRecord::EpochEnd { time, next, summary } is in bxcore and unused. messages already ignores it. This plan writes it.
• Session::open reads only 0.baseline.json and 0.jsonl. turns() and untrusted() look at that file only. next_call is recovered from that file only.
• run_turn always sends slots.main. Live turns stay on the session's live slot (main, from SessionStart). Only the summary call uses slots.background. Changing live turns to the session's slot is M5d-2.
• memoryd's indexer already walks every <epoch>.jsonl. Honcho's log_path is still sessions/<id>/0.jsonl.
• Runtime has no MemorySource. Compaction asks memoryd for a fresh block from channel.rs (and compact::finish_turn), so given turn tests that build a Runtime do not change.
• [compact] summarise is stored from this plan so a spec-complete loopd.toml loads; nothing in M5d-1 reads the list. M5d-2 opts in http_fetch.

Global constraints

• Roles depend on bxcore only. No unsafe, no async runtime, no new crate.
• No file the implementer reads holds a chat-template tag whole (T33). The carry heading is built from pieces and named in words.
• Every path a given test creates is removed by a guard (T44); every process a test starts has a deadline.
• The fail-closed state session-compact-failed has a runbook entry, added by the task that adds its pointer (T49).
• Each task gives its baseline and its gate count with what the count includes (T50, T53); the branch line is a check, never git switch (T51).
• Cargo.lock is in the git add line (T24). Gate timeout 600000 ms (T40).
• Commit trailer from the driver.

Tasks

Every task below has been written, replayed from its text and checked; the gate count is the passed total of cargo test --workspace plus one, for the ignored end_to_end test the gate runs on its own. Before task 01 it is 1456 (straylight, rustc 1.98.1 / clippy 0.1.98, 48a229ceae; master 3726ae5 after the M5d plans). End-state make gate on the reference is 1508.

#	Delivers	Crates	Gate after	Tests added	Mutations tried, caught
01	[compact] on loopd.toml: threshold 0.85, summary_chars 4000, summarise default ["http_fetch"] (stored, unused); Config::compact_prompt is compact.md beside system.md	loopd	1464	8	5, 5
02	Session::open loads the highest epoch that has both a baseline and a log, the carry from the previous EpochEnd, turns / untrusted / next_call over every epoch; persist writes the current epoch	loopd	1473	9	5, 4 (log_path after open does not move the already-open File)
03	messages_with_carry: epoch n > 0 is the new system message, then one ChatMessage::User (fixed heading plus carry, not a log record), then the new epoch's records; run_turn uses session.messages()	loopd	1478	5	2, 2
04	Honcho concatenates every <epoch>.jsonl in epoch order; turn_records of turn 3 is the epoch-1 user; send_turn reads the concatenated log	memoryd	1484	6	3, 3
05	loopd::compact: due, fallback_summary; starter config/compact.md	loopd	1490	6	3, 3
06	Summary call on the background slot, no tools, thinking on; tag-neutralised and capped; context-full uses the fallback; session-compact-failed pointer and runbook entry	loopd	1496	6	3, 3
07	Writing the next epoch: EpochEnd, fresh baseline from memoryd, SessionStart on the live slot; resume if EpochEnd is already the last record	loopd	1502	6	2, 2
08	After a turn: compact when due; on SessionFull, compact and retry the user message once; compaction holds the session busy	loopd	1508	6	2, 2 (finish_turn after drop(held) is wiring in channel.rs; the given tests call finish_turn directly)

After task 08, the design model runs the M5d-1 device check (spec 7) on straylight: a bxctl session writes a distinctive fact into memory.md, continues until compaction with [compact] threshold lowered, and the next turn's baseline contains the fact; recall of a phrase from epoch 0 still hits; TurnDone.untrusted stays true across the epoch change after a tainted recall.

Known limits

These are known, accepted for M5d-1, and not bugs to fix in a task:

• Live turns still use slots.main. SessionStart.slot is recorded and copied onto the new epoch; run_turn does not read it yet. Isolated sessions on the background slot are M5d-2.
• [compact] summarise is unused. Unknown tool names are not a config error until M5d-2.
• No idle path exists. Compaction is only called after a turn or on SessionFull. There is no timer to test.
• A crash after EpochEnd and before the next files exist reopens epoch 0 with EpochEnd at the end of the log. The next compact resumes without writing a second EpochEnd (task 07).
• A SessionFull retry logs the user message twice, once in the old epoch (Honcho skips it: no reply) and once in the new epoch. Turn numbers stay monotonic.

Before hand-over

Work through docs/plans/CHECKLIST.md with a reference in a scratch worktree with its own build directory, replayed from its task text; mutations; /tmp compared around a gate run; scratch files in the worktree's .state/; the reference moved to straylight as a bundle before the run (T18). Task files name their log row (T47).

Running it

The reference (m5d1-ref, tasks 01 to 08) is moved to straylight as a bundle and deleted here before the run (T18). Then, from the repository root, with the OpenCode TUI closed:

git switch -c m5d1 master
OPENCODE_MODEL=straylight/ornith-1.5-35b-a3b RESUME=1 \
  SLOTS_URL='http://straylight:11434/slots?model=ornith-1.5-35b-a3b' \
  tools/run-plan.sh docs/plans/M5d1 01

After task 08, the gate count is 1508. Then the device check above.