fix: eight things that stopped a second repository getting through the loop

[thedancingdeveloper]

Eight fixes from one exercise: importing rdpapp — a Rust/TypeScript product
with a 634 KB gateway file and a 272 KB browser entry point — and driving items
through the loop until they landed. Closes #150, #154, #156, #157, #159.

Based on fix/validator-rejects-valid-patches (#145), not main, because the
code these change only exists there.

Every one was found by watching a real item fail. Two items now go end to
end that could not before.

---

Context: what each role could see

1. The planner had never seen the repository (the most consequential).
Its entire job is naming the files an item will change, and it was given the
brief and nothing else — so it could only name a path the brief had already
quoted. The first item that described a surface rather than a path replied
"I do not have the repository tree or actual file paths" with an empty target
list, and the implementer then wrote a diff against a stylesheet it had not
been shown. It now gets the tracked paths, bounded by the same budget.

2. A target too large to supply was dropped silently (#150). The
implementer was asked to change a file it had never seen. Now the item stops
before that call is paid for — escalated / context_unavailable, blocked,
costing no attempt — and --context-budget / $HARNESS_CONTEXT_BUDGET exist,
because the budget was a constant with a context_policy nothing ever passed.

3. …but only a missing primary target is fatal. The first version of
that guard blocked whenever any named target did not fit, which on a
repository with two large files lost an item whose real target had been
supplied. The planner orders targets by importance; supporting files are now
named in the prompt with an instruction not to change them.

4. The reviewer got the diff and nothing else (#156). Two of three reasons
in a rejection were "the diff does not show whether…", which no diff can —
while the implementer had 700,000 characters of the same repository. It now
gets the touched files, and a file too large to include is named as absent.

Instructions: what each role was told

5. The reviewer was still told to reason about "the diff alone." Supplying
the files was only half of it — with 637,317 characters in hand including the
whole target file, it still wrote "I cannot verify there isn't another path…".

6. The implementer was not told which checks judge it (#154). A correct
change was refused by cargo fmt --all -- --check, a command it had never
seen.

7. A retry did not know why the last attempt failed (#157). Three of four
attempts at one item were refused by the same formatter, each unaware the
previous had been. requeue keeps last_error deliberately; nothing passed it
on. Not resumption — D11 is untouched, no prior diff is fed back, and the
prompt says so in as many words.

8. The reviewer rejected for things that were not the work (#159). It could
not confirm "no other file changed" (a diff is the whole change) or that the
tree type-checked (Checks: passed reads as the author's claim; the harness
ran them). And it overruled an explicit scope statement — verifying every
requirement, quoting the exclusion back, and rejecting because the wider
problem remained. The prompt now states both facts and gains its first "do not
reject if": scope is the task's to set.

---

Evidence

Two items completed plan → implement → apply → checks → commit → review → done, each on its own branch. The second is the direct test of #8 — same
model, same diff, third attempt:

I cannot verify from this change alone that main.tsx has been updated to
use them, but that was explicitly out of scope. — APPROVED

What these do not solve is #158: 8 of 23 model diffs did not apply, and the
rate tracks target file size (0/1 against 258 bytes, 8/22 against 272 KB and
634 KB). That is a protocol question under D10 and is deliberately untouched.

Also corrects the reason-kind list in schemas.py and USAGE.md, four kinds
behind outcomes.py, and documents what each role is shown.

Gates

All four from the repository root with TMPDIR on a fast volume: pytest
green, ruff check, ruff format --check, mypy (full-project strict) clean.

Fourteen new tests, each asserting behaviour rather than a flag: that the
implementer role is never called when its primary target is missing, that a
supporting one is named instead of fatal, that the planner gets paths and not
contents, that a project with no checks tells its reviewer it is the only gate.

🤖 Generated with Claude Code