boxmaker/docs/implementer-lessons.md

Implementer lessons

What reviews of the implementing model's work (Laguna S 2.1 through OpenCode) have taught us, as tips. It has two audiences: the implementer, and whoever writes its tasks. Every tip comes from a finding in docs/implementer-log.md. A tip without evidence does not belong here.

How it is used:

• After each milestone review, each finding becomes a row here or confirms an existing one.
• Short, general rules are promoted into AGENTS.md, which the implementer reads in every session and which must stay small. The "In AGENTS.md" column says which.
• "Seen again" is filled in at later reviews. A tip that does not stop its defect from coming back needs rewording or a test, not repetition.

Tips for the implementer

#	Tip	Evidence	In AGENTS.md	Seen again
I1	When a rule says "every" or "everywhere", finish by listing each place it could apply (every struct and enum in the file, every script, every branch) and check them one by one. An example in the task shows one place; the rule covers all of them.	M1 finding 1: deny_unknown_fields was put on both enums in task 06, where the task's example showed it, and on neither struct. In task 03, where the task said "works on structs and on enums", every type got it.	yes	M2a: no. All six config structs have it; none of the server-format structs do.
I2	A check must fail when it cannot do its job: missing input, unreadable file, a tool that errors. Never throw errors away (2>/dev/null, || true, an ignored Result).	M1 finding 4: all three gate scripts passed when crates/ did not exist, and hid find and awk errors.	yes	M2a: no recurrence.
I3	Report every problem found, not only the first.	M1 finding 8: check-lines.sh exits at the first long file.	yes	M2a: not exercised.
I4	Run the formatter before the gate, and let it decide ordering.	M1 finding 7: three first-gate failures were rustfmt reordering lib.rs.	yes	M2a: no fmt failures; the first-gate failures were clippy.
I5	Log every attempt, including one you abandon.	M1 review: the task 01 row mentions files "from a prior attempt" that has no row of its own.	yes	M2a: partly. Task 11's stopped row was overwritten by its later done row.
I6	Prefer a conversion that can fail (u32::try_from) to a cast that is only safe because of a check somewhere else (as u32).	M1 finding 6: bounded as casts in frame.rs. Correct today, but the safety depends on a line ten lines away.	no (already implied by the code rules)	M2a: no as casts at all.
I7	A read returns as soon as it has any data to give. It blocks only when it has none. Never keep reading to fill the caller's buffer.	M2a finding 2: the chunked reader delivered a whole stream at its end.	yes
I8	When forwarding between two connections, pass a close on in both directions. Do not rely on the client half-closing first.	M2a finding 1.	no (specific to inferproxy)
I9	Never end a turn by describing what you are about to do. Do it, then report.	Two Ornith turns in M2a ended with a "## Objective" plan and no tool call.	yes
I10	If a tool you were told to use does not exist, stop and say so. Do not invent a command in its place.	Laguna as coordinator ran opencodec, which does not exist, and then diagnosed its own typo.	yes

Tips for writing tasks

#	Tip	Evidence
T1	The implementer follows examples more reliably than prose. For a rule that applies in several kinds of place, show an example of each kind, or say in so many words that it applies to structs and enums alike.	Same evidence as I1. The one task whose note covered both kinds had no miss.
T2	A test for a property that must hold "everywhere" should walk the data, not sample it.	M1 finding 1 got through because the given tests checked unknown fields in a few hand-picked places. tests/strict.rs now adds an unknown key to every object of every fixture.
T3	When a rule depends on a file format, list every way the format can write the same thing.	M1 finding 3: the task said "however it is written" and named three spellings; TOML has a fourth, the [dependencies.name] table, and the dependency check passed it.
T4	If formatting or serialising a value can fail, make the value impossible to construct. Check what the library does at the edges of its range before writing the signature.	M1 finding 2: Timestamp accepted any u64; humantime cannot format past year 9999 and to_string() panics. The reference implementation had the same bug, so the given tests could not catch it.
T5	A reference implementation proves the tests can pass. It does not prove they are complete: reference and tests share the author's blind spots. Probe the finished work from outside with inputs the tests do not contain.	Findings 2 and 3 were in the reference too. Both were found by probing, not by the gate.
T6	Do not tell the implementer how to lay out what a formatter controls.	M1 finding 7.
T7	A follow-up task works well as: what the reviewer observed, which part was the task's fault, a failing test to copy in, and the new rules.	Tasks 08 and 09 each passed the gate on the first run, and 09 handled six forms its self-test did not contain.
T8	Vet a dependency by testing it against an independent implementation before naming it in a spec, whoever wrote it. Include the boundary cases of its algorithm.	The owner's emsha 1.0.3 passed its own tests and hashed every message of length 63 mod 64 wrongly. A 90-case differential run against sha256sum found it in minutes. Same shape as T5: an author's tests share the author's blind spots.
T9	State the contract of a standard trait the implementer must honour, in the task, even when it seems obvious.	M2a finding 2: task 04 said what Body reads but not when read must return. Two implementations, the implementer's and the reference, both got it wrong in different degrees.
T10	Test when streamed data is delivered, not only what. A test that compares final bytes cannot see buffering.	M2a finding 2 passed 15 tests that only compared bytes.
T11	Make the fake client behave like the real one. If the real client never half-closes, no test client may half-close either.	M2a finding 1: all test clients half-closed; loopd does not.
T12	A weak model should not coordinate other sessions. Use the shell driver (tools/run-plan.sh), whose checks are code.	The opencodec episode.
T13	Set the sampling defaults on the server for coding agents (temp, top-p, top-k, reasoning-budget). OpenCode sends none, so the server's default temperature of 1.0 applies, and thinking has no cap.	Five Ornith turns ran 16k tokens of thinking to the output limit and produced nothing.

What worked and should be kept

• Byte-exact fixtures, compared in both directions. No wire-format defect reached review.
• Tests and fixtures copied in task by task and never edited. All 23, then 25, files were byte-identical at both reviews.
• A step that makes the implementer prove a test has teeth (task 07: make new public, watch the compile_fail doctest fail, change it back). It was done and logged.
• "Stop and report" instead of improvising. There were no invented dependencies, no #[allow], no edited tests and no touched design documents across nine tasks.
• One task, one fresh session, one commit, with only listed paths staged. The history reads as the plan.

What we have seen of the implementers so far

Nine tasks, all in M1, all small and pinned by tests, so treat these as first impressions.

• Follows process rules exactly: branch, staging, trailer, log, protected files.
• Writes idiomatic Rust without panics or unsafe shortcuts, and adds sensible things the task did not ask for (FrameError::source).
• Misses the unshown half of a general rule (I1).
• In shell, reaches for defensive habits that hide failures (I2).
• Given a clear defect report and a failing test, fixes the defect minimally and generalises.
• Gate on the first run: 6 of 9 tasks. Wall time per task, including the owner's turnaround: about 10 to 55 minutes.

M2a added two more models. Ornith-1.5-35B-A3B did nine tasks: it found and fixed a real bug in its own chunked reader, wrote module docs everywhere, and stopped correctly when a prerequisite was missing; its two failure modes were runaway thinking to the output limit and ending a turn with a plan instead of a tool call. GLM-5.3 did one task, the assembler, correctly and quickly, with the best comments of the milestone. Laguna did three and a half tasks in M2a with more nudging than in M1, and failed as a coordinator. First-gate pass rate for the milestone: 8 of 13.