[solvyxtech]

Follow-up to the PR #51 re-review, non-blocking item 2.

Bug. The engine's shown map (which lines of a file the model has been shown) was keyed by the path as spelled. read_file dur.py, then a successful edit_file ./dur.py, then read_file dur.py: the edit deleted the ./dur.py key, dur.py stayed marked as shown, and the re-read was answered with the "already been shown … nothing has changed since" pointer, while the transcript (which strips ./) had already elided the old copy. The model held no copy of the file it had just edited.

Fix. canonPath in src/transcript.ts is the one path normaliser. Transcript elision uses it (where it replaces the two inline ^./ strips), and so does the engine at every place the shown map is read, set or deleted. It also folds sub/../dur.py, and with the workspace root, absolute paths inside it. Paths outside the workspace stay absolute. No other tool (bash, rename, delete) touches shown.

Repro, final tool results on the wire:

• before: [molt: you have already been shown lines 1-2 of dur.py. Scroll up rather than reading it again — nothing has changed since. …]
• after: def parse(s):\n return 5400 # NEW_MARK

Tests. test/lean-fixes.test.ts has engine-level regressions for dur.py/./dur.py both ways, an absolute path and sub/../dur.py. All four failed before the fix. There are also canonPath unit cases and a transcript elision case.

Verified (~/.local/bin off PATH):

• npm run check: 2439 tests, 2434 pass, 0 fail, 5 todo; e2e PASS, self-check PASS.
• lean-budget: identical to main on all 7 scenarios (re-read 44,725 / 348,643 / 9).
• adversarial fixtures: 47 pass, 0 fail, 5 todo.

[Solvyx]

Grok review of e0762f3, comment only, not merging.

Verdict: no blockers, fine to merge once CI is green (test-linux and typecheck-and-test were still pending when I looked; privsep-linux passed). The fix does what it says for the reported case. I checked out the head, ran typecheck and test/lean-fixes.test.ts locally (25/25 pass), and confirmed the engine and transcript can no longer split the way #51's item 2 described. canonPath-with-root depends only on posix.normalize of the input, so whenever the transcript treats a write and a read as one file, the engine does too. That means the transcript can't elide a copy the engine still calls "shown".

Pairing rules: clean. No subscription backends, no competitor names in copy, nothing touching fine-tune data, and it's on a claude/ branch.

Blocking

None.

Non-blocking

1. The transcript doesn't fold absolute paths, but the changelog and docstring say it does. (src/transcript.ts:636, :640; CHANGELOG.md:159; confidence high, impact low.) The engine calls canonPath(path, this.cwd), while the transcript calls canonPath(path) with no root. Repro: read_file /ws/dur.py, then edit_file dur.py, then elideSupersededReads() elides 0, so the stale pre-edit copy stays in context. The engine side is correct (both spellings map to dur.py), so this is wasted tokens rather than a missing copy. But the CHANGELOG line ("the engine and the transcript now key files through one normaliser … an absolute path inside the workspace as the same file") and the canonPath doc ("have to agree on it") overstate it, and the absolute-path test at test/lean-fixes.test.ts:329 only covers the engine. Fix: give Transcript (or elideSupersededReads) the workspace root, or narrow the wording.

2. On Windows, absolute paths still get the wrong key, so the original false "nothing has changed" comes back. (src/transcript.ts:44-52; confidence medium. release.yml ships a win build, but CI only tests on Linux.) path.relative("C:\\ws", "C:\\ws\\src\\dur.py") returns src\dur.py, and posix.normalize leaves the backslash, so the key doesn't match a read of src/dur.py. Scenario: read src/dur.py, then edit C:\ws\src\dur.py (which deletes the wrong key), then re-read src/dur.py. The re-read gets "already been shown … nothing has changed since", and since the transcript keys differ too, the stale copy is the only one in context. The !rel.startsWith("../") guard at :49 also misses ..\ there. This isn't a regression (main was already wrong here), but the PR claims the absolute-path case is fixed. Suggest replacing \ with / after relative() when sep === "\\", plus a unit case built on path.win32.

3. Follow-up idea: make "nothing has changed since" true by checking the file, not the tool name. (src/engine.ts:5834; confidence high, pre-existing, out of scope.) As the PR says, only write_file/edit_file clear shown. So read dur.py, then bash: sed -i … dur.py (or a test run that rewrites it, or an apply_patch-style shell edit), then re-read, and the model is still told nothing changed while it holds the old copy. Storing a size+mtime (or a short content hash) with each shown entry and dropping its coverage when that changes would cover bash writes, renames, symlinked-cwd spellings (/var vs /private/var on macOS) and case-only spellings on case-insensitive filesystems all at once. It also fits the project's line about proving things against disk. Worth a separate issue rather than growing this PR.

Nits: none worth holding for.