Make compile_feedback work-item aware: blast-radius scope + immutable-base baseline + lint budget #294

Merged

toasterson merged 5 commits from claude/wi-019fa83d-compile-feedback-lints-the-whole-workspa into main 2026-08-08 13:46:16 +00:00

Conversation 1 Commits 5 Files changed 1 +1235 -144

toasterson commented 2026-08-06 18:24:57 +00:00

Owner

Copy link

Anima work item 019fa83d-523b-7303-b962-13d10c9ae483.

Measured 2026-07-28

WI-230 (Wave R1 — Config spine) ran four rounds on tecton and produced zero
config-spine work. Every round opened the same way: run compile_feedback, get
workspace-wide clippy warnings in crates/grug-data and benchmarks/harness-core,
start fixing them, hit the turn watchdog, end. Rounds 1–4 pushed 48 lines of that
cleanup under the title "Wave R1 — Config spine", and the reviewer rejected PR #246
for containing "work from MANY work items".

The warnings are real but trivial and unrelated: collapsible if let → let-chains,
write! with a trailing newline → writeln!.

Why the akh sees what nobody else does

.forgejo/workflows/ci.yml runs cargo check -p seshat — one crate. Its own
comment explains why: broader test/clippy/release-build OOM-kills rustc on the
Solstice runner. So workspace-wide clippy has been rotting unchecked while main
stays green, and compile_feedback is the only thing that looks at it.

The akh then reasonably concludes the tree is dirty and cleans it — on its own
budget, in someone else's crates, every single session.

Ask

1. Scope compile_feedback to the work item's blast radius. Default it to the
   crates the session actually touches (worktree diff, or the package(s) owning the
   edited paths); --workspace becomes opt-in via the tool's parameters.
2. Separate "my change broke this" from "this was already broken". Capture a
   baseline at session start; report pre-existing diagnostics as context ("the tree
   arrived with N warnings in crates you are not touching") rather than as work.
   The akh must not be able to mistake ambient rot for its assignment.
3. Consider a lint budget: if pre-existing warnings outside the blast radius exceed
   a threshold, say so once and move on.

Related

• akh-medu PR #274 — the cleanup itself, extracted so it lands once on main.
• Anima 019fa7c6-ec4b-73e1-8b43-dc4deeca8373 — the runner must not commit / push /
  PR a turn that only repaired the base repo.
• The turn watchdog half of this (300s < glm-5.2's ~5-minute responses) was mitigated
  the same day by setting AKH_TURN_IDLE_TIMEOUT_SECS=900 on the akhomed container.

Anima work item `019fa83d-523b-7303-b962-13d10c9ae483`. ## Measured 2026-07-28 WI-230 (Wave R1 — Config spine) ran four rounds on tecton and produced **zero** config-spine work. Every round opened the same way: run `compile_feedback`, get workspace-wide clippy warnings in `crates/grug-data` and `benchmarks/harness-core`, start fixing them, hit the turn watchdog, end. Rounds 1–4 pushed 48 lines of that cleanup under the title "Wave R1 — Config spine", and the reviewer rejected PR #246 for containing "work from MANY work items". The warnings are real but trivial and unrelated: collapsible `if let` → let-chains, `write!` with a trailing newline → `writeln!`. ## Why the akh sees what nobody else does `.forgejo/workflows/ci.yml` runs **`cargo check -p seshat`** — one crate. Its own comment explains why: broader test/clippy/release-build OOM-kills rustc on the Solstice runner. So workspace-wide clippy has been rotting unchecked while main stays green, and `compile_feedback` is the only thing that looks at it. The akh then reasonably concludes the tree is dirty and cleans it — on its own budget, in someone else's crates, every single session. ## Ask 1. **Scope `compile_feedback` to the work item's blast radius.** Default it to the crates the session actually touches (worktree diff, or the package(s) owning the edited paths); `--workspace` becomes opt-in via the tool's parameters. 2. **Separate "my change broke this" from "this was already broken".** Capture a baseline at session start; report pre-existing diagnostics as context ("the tree arrived with N warnings in crates you are not touching") rather than as work. The akh must not be able to mistake ambient rot for its assignment. 3. Consider a lint budget: if pre-existing warnings outside the blast radius exceed a threshold, say so once and move on. ## Related - akh-medu PR #274 — the cleanup itself, extracted so it lands once on main. - Anima `019fa7c6-ec4b-73e1-8b43-dc4deeca8373` — the runner must not commit / push / PR a turn that only repaired the base repo. - The turn watchdog half of this (300s < glm-5.2's ~5-minute responses) was mitigated the same day by setting `AKH_TURN_IDLE_TIMEOUT_SECS=900` on the akhomed container.

toasterson added 1 commit 2026-08-06 18:24:59 +00:00

compile_feedback lints the whole workspace, so every session burns its turns cleaning crates the work item never touches ... da5ec5a55c

Anima implementation session.

toasterson added 1 commit 2026-08-06 19:02:23 +00:00

compile_feedback lints the whole workspace, so every session burns its turns cleaning crates the work item never touches ... 2071f0d31a

Anima implementation session.

toasterson force-pushed claude/wi-019fa83d-compile-feedback-lints-the-whole-workspa from 2071f0d31a to 296ae33f7b 2026-08-06 20:46:53 +00:00 Compare

toasterson reviewed 2026-08-07 16:48:59 +00:00

toasterson left a comment

Copy link

Review verdict: BLOCK — the branch does not compile.

cargo check --features server at 296ae33f in a clean worktree fails with two hard errors in the code this PR adds:

• src/agent/tools/compile_feedback.rs:172 — E0277: .unwrap_or_default() on Option<std::process::Output>, but std::process::Output does not implement Default.
• src/agent/tools/compile_feedback.rs:409 — E0283: redundant .into() on a &'static str passed to ToolOutput::ok, which takes impl Into<String>.

Note that CI green is not a signal on this repo: .solstice/job.sh runs only cargo check -p seshat and never builds the root akh-medu crate where this file lives.

Beyond the build break, six substantive defects remain — silent fail-closed on git-diff failure, a blast-radius path check hardcoded to crates/<name> that misses benchmarks/harness-core (one of the two crates this work item exists to fix), a baseline dedup key that ignores file and line, no automatic baseline capture at all, and zero tests over ~200 lines of new logic. Full detail is in the work item description.

**Review verdict: BLOCK — the branch does not compile.** `cargo check --features server` at `296ae33f` in a clean worktree fails with two hard errors in the code this PR adds: - `src/agent/tools/compile_feedback.rs:172` — E0277: `.unwrap_or_default()` on `Option<std::process::Output>`, but `std::process::Output` does not implement `Default`. - `src/agent/tools/compile_feedback.rs:409` — E0283: redundant `.into()` on a `&'static str` passed to `ToolOutput::ok`, which takes `impl Into<String>`. Note that CI green is not a signal on this repo: `.solstice/job.sh` runs only `cargo check -p seshat` and never builds the root `akh-medu` crate where this file lives. Beyond the build break, six substantive defects remain — silent fail-closed on git-diff failure, a blast-radius path check hardcoded to `crates/<name>` that misses `benchmarks/harness-core` (one of the two crates this work item exists to fix), a baseline dedup key that ignores file and line, no automatic baseline capture at all, and zero tests over ~200 lines of new logic. Full detail is in the work item description.

toasterson referenced this pull request 2026-08-07 18:40:19 +00:00

CI: build the akh-medu crate, and stop running every job twice #296

toasterson force-pushed claude/wi-019fa83d-compile-feedback-lints-the-whole-workspa from 296ae33f7b to b1c52a51f6 2026-08-07 18:45:25 +00:00 Compare

toasterson added 1 commit 2026-08-07 21:34:38 +00:00

compile_feedback lints the whole workspace, so every session burns its turns cleaning crates the work item never touches ... 6d719f755e

Anima implementation session.

toasterson added 1 commit 2026-08-08 11:43:20 +00:00

Make compile feedback work-item aware 1eba90408d

toasterson changed title from WIP: compile_feedback lints the whole workspace, so every session burns its turns cleaning crates the work item never touches to Make compile_feedback work-item aware: blast-radius scope + immutable-base baseline + lint budget 2026-08-08 13:20:38 +00:00

toasterson added 1 commit 2026-08-08 13:26:39 +00:00

ci: re-trigger Solstice on 1eba904 (prior run died on archibald saturation) ... 91fa5a3a84

All six jobs on 1eba904 failed at ~50-58min, including publish-chart, which
only touches charts/ and was untouched by this branch and green on the parent
commit 6d719f7. archibald was carrying load average 52 on 16 cores with 11GB
swapped at the time. Infra, not code.

Verified locally on this tree: cargo check -p akh-medu clean,
cargo test --lib compile_feedback 20/20 passed.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

toasterson merged commit 8c365ac54b into main 2026-08-08 13:46:16 +00:00

toasterson referenced this pull request from a commit 2026-08-08 13:46:19 +00:00

Merge pull request 'Make compile_feedback work-item aware: blast-radius scope + immutable-base baseline + lint budget' (#294) from claude/wi-019fa83d-compile-feedback-lints-the-whole-workspa into main

Sign in to join this conversation.

Reviewers

No reviewers

Labels

Clear labels

No items

No labels

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

1 participant

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference

toasterson/akh-medu!294

No description provided.