CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

Current state

Boxmaker is a sovereign personal agent harness written in Rust. Work proceeds one milestone at a time (M0 to M7, table in docs/milestones.md). M0 (measurements), M1 (workspace, proto, gate, Decision), M2a (the inference path: inferproxy, loopd's HTTP, SSE and llama client, the startup self-test) and M2b (sessions, the turn loop, loopd serve, bxctl chat) are done. M3 is split: M3a (brokerd's decision path, spec docs/specs/2026-09-18-m3a-decision-path.md) is done and merged, reviewed in docs/implementer-log.md; its open findings (17, 18, 20) are listed there; 14 was closed by M3b task 02. M3b (the container runner and the four tools, docs/specs/2026-09-22-m3b-runner.md) is done and merged: tool containers run from the image deploy/tools-image.nix builds, checked on straylight. M4 is split: M4a (gatewayd, conversations over Mattermost, spec docs/specs/2026-09-23-m4a-gateway.md) is done and merged, reviewed in docs/implementer-log.md and checked end to end against the owner's Mattermost; no finding is open. M4b (approvals over Mattermost, spec docs/specs/2026-09-24-m4b-approvals.md, 18 tasks) is done and merged, reviewed in docs/implementer-log.md and checked end to end against the owner's Mattermost, including an approval asked by a non-approver; no finding is open. M5 (memory, then the scheduler) is split into M5a, M5b and M5c by P17 to P21 (applied 2026-09-24). M5a (local memory, memoryd, spec docs/specs/2026-09-24-m5a-local-memory.md, 22 tasks) is done and merged, reviewed in docs/implementer-log.md and checked on straylight against the owner's Mattermost; the device check found that a recalled tainted note left a session trusted, fixed by task 21. P24 split the rest of M5. M5b (hardening: the sanitizer, tag neutralisation in loopd, images and untrusted links as code in gatewayd, the invalid-arguments reason, bxcore::secrets, one instance per daemon; spec docs/specs/2026-09-26-m5b-hardening.md, 14 tasks) is done and merged, reviewed in docs/implementer-log.md and checked on straylight against the owner's Mattermost; the device check found that the self-test overran inferproxy's burst (fixed in review). M5c is Honcho, through the owner's roncho (v1.0.0 on crates.io; docs/specs/2026-09-25-roncho-requirements.md and docs/specs/2026-09-27-m5c-honcho.md). M5c-1 (the write path) is done and merged, reviewed in docs/implementer-log.md. M5c-2 (the read path: recall over the fleet and ask_memory, spec sections 3, 8, 9, 10 and 13) is done, reviewed in docs/implementer-log.md and checked on straylight against rift; no finding is open. M5d (scheduler and compaction, spec docs/specs/2026-10-03-m5d-scheduler.md) is split: M5d-1 (epochs and compaction) is done, reviewed in docs/implementer-log.md and checked on straylight (spec 7); no finding is open. M5d-2 (heartbeat, cron, ToolSummary, live-slot isolation) is done and on master, reviewed in docs/implementer-log.md; finding 1 closed by task 11; spec 7 device-checked on straylight (heartbeat on slot 2 left a following slot-0 turn's cache_n at 1398 of 1399; jobs/tier.toml promoted only a trusted note). M6 (subagents and cloud consult) is approved (docs/specs/2026-10-04-m6-subagents-consult.md). M6a (subagent tool, isolated child on slots.background, scheduler skip, brokerd taint inherit; spec docs/specs/2026-10-04-m6-subagents-consult.md sections 3, 5, 6, 7 M6a and 8; plan docs/plans/M6a/) is done, reviewed in docs/implementer-log.md; no finding is open; spec 7 device-checked on straylight (child sub-m6a-live-2 on slot 2; following slot-0 turn cache_n 1787 of 1788). M6b (consult behind brokerd, allowlisted loopback HTTP, Decision body; spec docs/specs/2026-10-04-m6-subagents-consult.md sections 4, 5, 6, 7 M6b and 8; plan docs/plans/M6b/) is done, reviewed in docs/implementer-log.md; no finding is open; spec 7 device-checked on straylight (clean consult audit body matched the captured loopback POST; a private untrusted recall still consulted under max_taint = "private", matching spec 4.1's class gate). P22 (proposed, not applied) makes Boxmaker a fleet of independent agents, one per host, deployed with Nix in a last milestone, M8; flake.nix builds every binary, the tool image and a dev shell. Work happens on titan (macOS, no Nix); build and run on straylight, where a clone is built with nix build. The owner's Honcho client roncho is at ~/src/roncho. Branches land on master by rebase and fast-forward, never a merge commit. The owner's Mattermost settings and the bot's token are in mattermost.env at the repository root (on straylight, ~/.config/boxmaker/mattermost.env): ignored by git, never committed, and the token is never printed. Straylight now serves Ornith as four slots over one 262,144-token pool; see docs/inference-contract.md, "Deployment change, 2026-09-20", before relying on cache behaviour. docs/runbook.md has an entry for every fail-closed state; any new one needs an entry and a pointer in its message. Check docs/implementer-log.md for what is in flight.

docs/design.md is the binding design brief. If it looks wrong or conflicts with a measurement, stop and say so. Changes to it land as their own commit and are recorded in docs/decisions.md, which also lists proposed changes that are not yet applied.
docs/inference-contract.md holds the M0 measurements from straylight. Where it and the brief disagree, the measurements are newer.
docs/specs/ holds design specs; docs/plans/M<n>/ holds the task files Ornith works from, and files/ under it holds the tests and fixtures the tasks copy in. AGENTS.md holds the implementer's standing rules; its code rules apply to any code written here too.
docs/implementer-log.md is the implementer's own record, one row per task. Reviews go at the bottom.
docs/implementer-lessons.md turns review findings into tips, for Ornith and for writing its tasks. Every milestone review updates it, and short general rules are promoted into AGENTS.md. Do not edit the repo while Ornith has a task in flight: its tasks require a clean working tree.
spike/ is throwaway measurement code, not harness code.

Roles: implementation is done by Ornith-1.5-35B-A3B (served by straylight) through OpenCode on this machine, which reads AGENTS.md. Builds are deployed to straylight afterwards. Design, specs, measurement and review are done here. Write plans for Ornith as small closed tasks: exact paths, given type signatures, verified crate APIs, tests specified first, and "stop and report" instead of judgement calls. The model the harness serves is Ornith-1.5-35B-A3B.

The inference server is shared with other sessions. Before using Ornith slot 1 or sending unpinned requests, check GET /slots?model=ornith-1.5-35b-a3b so you do not evict someone's cache.

Commands

make gate runs offline: cargo fmt --check, clippy with warnings denied, cargo test, cargo-deny, and a check that fails on any source file over 500 lines. Run it before calling any work done, and report the exit status and last lines.
Single test: cargo test -p <crate> <test_name>.
make audit runs cargo deny check advisories; it needs the network.
nix build (on a host with Nix, such as straylight) builds every binary; nix build .#tools-image the tool image; nix develop gives the toolchain. See docs/dependencies.md.
make verify-device (from M2a task 13) runs the ignored tests in crates/loopd/tests/device.rs against straylight through a private inferproxy. It uses Ornith slot 0. tools/check-m3a-device.sh runs brokerd, loopd and an approval end to end against it.
bxctl is the owner CLI: bxctl chat talks to a running loopd serve --config <path>; bxctl approvals, approve, refuse, grants check and audit verify talk to brokerd serve (or read the audit log directly); bxctl reindex comes in M5a. gatewayd serve --config <path> is the Mattermost channel (M4a).

Architecture in brief

Separate binaries in one Cargo workspace. Each role holds as little authority as possible:

loopd owns sessions and prompt assembly. It has no credentials and no network. Its only I/O is Unix sockets to gatewayd, brokerd, inferproxy and (from M5a) memoryd.
memoryd (M5) owns memory: memory/soul.md, memory.md, user.md (identity), notes.md (experience), the FTS5 index, and from M5c the sync to the fleet's Honcho. It runs memory tool calls brokerd has decided, and reaches Honcho and nothing else.
brokerd is the only place authority lives. It reads owner-written grants (it cannot write them), runs each approved tool call in a fresh rootless container, and writes a hash-chained JSONL audit log.
gatewayd is the Mattermost channel. Outbound only, no listening port; approvals arrive as !approve/!deny replies over the WebSocket.
inferproxy is a ~100-line byte forwarder from a Unix socket to the shared llama-server router on the host. It is required, because loopd has no network.
bxcore (formerly proto) holds shared types, and from M5b the secret store and the sanitizer. toolkit holds tool container entrypoints.

Structural rules that span crates:

No crate depends on another role's crate. Crates depend only on bxcore.
Authority is encoded in types. A tool can't run without a Decision. It is defined in brokerd with a private field and no Deserialize, so only the policy module can construct one (covered by a compile-fail test). bxcore carries a plain DecisionRecord.
No source file over 500 lines.
Files are the source of truth (sessions/, memory/, grants/, audit/). SQLite is only for rebuildable indexes and queues.

Inference contract (the constraint most likely to be broken by accident)

Prompt processing on straylight is slow, and the hybrid-attention model can't partially rewind its KV cache, so any change to an earlier byte of the prompt forces an expensive full re-read.

Each turn's request must be a strict extension of the previous one. Volatile content (time, heartbeat notes, memory refreshes, recalled memory) goes only in the newest message, never in the system prompt or earlier history.
Assistant messages are logged and replayed exactly as the server returned them, including reasoning_content and tool_calls.
The tools array is fixed per epoch; changing it re-reads the whole prompt. Other tools are found with find_tool and called through the call_tool meta-tool. Never prompt the model to call an undeclared tool: the server's grammar forces the call into a declared one.
Tool results are size-capped when first appended and never trimmed later.
The baseline (system prompt, tool schemas, the three memory files) stays at 3,000 tokens or less, measured with the server's tokenizer.
Compaction runs after a turn at or above 85% of n_ctx (configurable), and starts a new epoch (sessions/<id>/<epoch>.jsonl). Old logs are kept. A quiet session stays on its current epoch.
Every request carries id_slot, and a session never changes slot. The server is shared, so slots are not reserved: a cold cache is a normal event that loopd logs, never an error.
Streaming always, with return_progress: true. Timeouts are "no bytes for N seconds", never total deadlines, plus a longer limit for waiting on a busy slot or a model load.

Working rules from the brief

Verify every external crate API on docs.rs, and every llama-server or Mattermost request field against primary docs, before use. Crate names and server parameters in the brief come from memory and aren't verified. If you can't fetch the docs, say so; don't guess.
Keep dependencies few. Justify each in docs/dependencies.md. Any outbound call must be listed in docs/egress.md. No telemetry and no update checks.
If a feature isn't in the brief, propose it; don't build it. Stay inside the current milestone's scope.
Write tests first. Make one logical change per commit. Never commit runtime data, secrets, or spike output that contains conversation content.

12 KiB Raw Blame History