school-core Last-Mile — Slice Plan & Dispatch Ledger

school-core Last-Mile — Slice Plan & Dispatch Ledger

Date: 2026-08-26 · Director: Brandon (via @sisyphus) · Status: dispatched

Rule zero (bead 9rn)

NO agent works in ~/school-core. Each gets an isolated clone. Live tree is read-only for the director. Violating this is how the repo ended in detached HEAD earlier today.

WorkerCloneSlice
@student-ci/tmp/wk-ciS1 — baseline green
@student-whymage/tmp/wk-whymageS2 — evidence comparator (diagnosis only)
@student-coder/tmp/wk-coderS3 — artifact hygiene
@student-executor/tmp/wk-execS4 — rebase #67 + re-measure

Dependency order (why the slices are sliced this way)

S1 baseline green ──┐
                    ├──> S4 rebase #67 & re-measure ──> verdict on #67
S3 artifact hygiene ┘
S2 evidence comparator (independent; unblocks the real discriminator)

S1 is FIRST because "green CI on real work" is unreachable while main itself is red (11 failures). S4 depends on S1+S3 because a rebase judged against a red baseline and a fence-polluted artifact tells you nothing.

S1 — Baseline green · @student-ci · bead school-core-fhm · P1

Clone: /tmp/wk-ci (branch main @ f82a6e7)

Measured starting state: 11 failed / 3 passed (targeted subset, py3.9 via /usr/bin/python3 — the Homebrew python3 has no pytest).

Known cause for the test_review_build_gate cluster: director.py:485

repo_path = _resolve_repo_path(repo, explicit_path=repo_path)
TypeError: <lambda>() got an unexpected keyword argument 'explicit_path'

A test double patches _resolve_repo_path with a lambda lacking that kwarg.

Contract

  1. Count CAUSES before fixing. 11 symptoms ≠ 1 cause. Classify each of: test_ce_router::test_route_decision_logs_to_bookbag, test_issue_bridge::TestCrewDispatchPath::{test_timeout_falls_back_direct,test_per_cycle_cap_falls_back_direct}, test_orca_executor_repo_path.py (2), test_review_build_gate.py (6).
  2. TDD: prove RED for the right reason before GREEN.
  3. Fix the DOUBLE or the CALLER — whichever is wrong. If explicit_path is a legitimate new param, the doubles are stale; if not, director.py is wrong. State which and why.
  4. Guard check first: grep test_n*, *_guard, *_invariant for anything asserting on _resolve_repo_path's signature.

Done when: targeted subset is 14/14 green in /tmp/wk-ci, staged diff reported, git diff --stat pasted, zero new failures vs the 11-failure baseline. Gate: no commit, no push.

S2 — Evidence comparator · @student-whymage · beads qpo + a9s · P1

Clone: /tmp/wk-whymage · DIAGNOSIS ONLY, no code

Measured: 125 status files in ~/.hermes/school-core-fm-config/state/ (82 working:, 24 done:, 14 failed:, 4 resolved:, 1 malformed). 87 cited SHA occurrences, 60 unique. 9 resolve in either ~/school-core or ~/sound-royale-ny; 51 resolve nowhere. Of the 9: 6 authored by Brandon (sound-royale-ny), 2 github-actions[bot], 1 fly-io[bot]. 0 crew-authored. git ls-remote origin 'refs/heads/fm/*' → 0 refs. And 0 of 125 files carry a project= field.

Question (why-tree to mechanism): why does _artifact_identity honor a commit= without validating (a) reachability and (b) authorship in the worktree?

Contract

  1. Trace to file:line. Name the checkout on every claim.
  2. This is a COMPARATOR case, not a vocabulary gap — both second sources (git cat-file -t, %an) were on disk the whole time. Do not propose a framework; propose the comparison.
  3. Reverse-tree the blast radius: what currently depends on done: being trusted?
  4. Separately: where SHOULD repo attribution be written so clause 3 of the discriminator becomes implementable? (@lucas suggests joining .status→.meta rather than adding a field — evaluate both.)

Done when: mechanism named at file:line, a 3-line comparator proposed (NOT a registry — a registry whose evidence source is weak passes forever while lying), and blast radius listed.

S3 — Artifact hygiene · @student-coder · beads 4ti + 03m · P2

Clone: /tmp/wk-coder

Two defects, both CONFIRMED:

  • pr_creator.py:466-468 writes school-output/<domain>/<num>/output.py with NO fence stripping. On the pr67 branch that file's line 1 is a literal ```python → SyntaxError.
  • .github/workflows/ci.yml:40 runs compileall -q . over the whole repo, including generated artifacts, so one bad artifact reds the entire matrix before a single test runs.

Verified exit codes (bare — never pipe, a pipe reports the LAST element's status):

compileall -q .                        -> 1
compileall -q -x "^school-output/" .   -> 1   # anchor bug: walked path is ./school-output/...
compileall -q -x "school-output" .     -> 0   # correct

Contract

  1. Producer: write output.md, strip outer fences (^```[a-zA-Z0-9_-]*\n and \n```$).
  2. Guard tests veto a naive rename: tests/test_pr_creator.py asserts on the .py path at :190, :221, :289, :311, :368. All five must move together.
  3. CI: ci.yml:40 → -x "school-output" (no ^).
  4. project_verify.yaml:21 uses compileall -q *.py (root glob) — unaffected, leave it.
  5. State what this does NOT fix.

Done when: both diffs staged, suite green in /tmp/wk-coder, compileall exit 0 on the pr67 branch there.

S4 — Rebase #67 & re-measure · @student-executor · bead 75o · P2

Clone: /tmp/wk-exec · BLOCKED until S1 and S3 land

Retraction context — read before touching this. The director previously reported #67 as "697 deletions incl. _STATUS_RE hardening." That was a MERGE-BASE ERROR and is RETRACTED:

merge-base(main, pr67) = a0b7646 (2026-08-17); main is 51 commits ahead
git diff main..HEAD    -> 35 ins / 697 del   &lt;- artifact of a stale branch
git diff a0b7646..HEAD -> 5 ins / 0 del      &lt;- truth (hello.txt + output.py)

PR #67 adds five lines and deletes nothing.

Contract

  1. Rebase pr67 onto current main in /tmp/wk-exec only. Report exact commands and exit codes.
  2. Re-measure tests/test_crew_dispatch.py::test_happy_path_reads_report_and_tears_down — it passes on main, fails on the stale branch. Hypothesis: STALENESS (old source vs current tests), not destruction. Confirm or refute.
  3. Always diff against merge-base: git diff $(git merge-base main HEAD)..HEAD.
  4. Deliver a verdict: mergeable / needs-work / genuinely-regressive, with the measurement behind it.

Done when: rebased locally, test re-measured both sides, verdict stated. Gate: no push, no force-push, no PR state change. Closing #67 would destroy the only end-to-end pipeline artifact that exists — the director decides.

Standing constraints (every slice)

  • One class per bead. No bundling.
  • Name the checkout on every claim (/tmp/wk-* + commit SHA).
  • CONFIRMED vs UNPROVEN labels, always explicit.
  • Tell the director if any premise in the brief is wrong. Premises here have been wrong four times today; two agents and the director all retracted.
  • HUMAN GATE: no commits, pushes, merges, or PR-state changes. Stage and report.
  • /usr/bin/python3 for pytest. Full suite ≈16 min — run it in background.
  • Never pipe a build/test command through head/tail; it masks the exit code.

Not delegated (director keeps)

  • Closing/merging anything.
  • The WG post and job-search sends.
  • Deciding #67's fate after S4 reports.
Built with LogoFlowershow