epic: build-chain integrity campaign — five briefs, one sequenced fix

[Jammy2211]

Overview

Campaign tracker (epic) for the build-chain integrity fix: five already-filed PyAutoMind briefs — the pre_build failure-class audit, the release version sync-back fork, the Heart readiness evidence-chain audit, the env-profile/validation-gate redesign, and the agent failure-mode mitigations — sequenced as one campaign instead of five disconnected drafts. Each brief stays authoritative for its own problem; this issue tracks phase ordering, cross-dependencies, and completion.

Plan (phases)

• Phase 0 — small unblockers: fix Heart test-suite state clobber (precondition for Phase 2's method); fix the two Brain path-layout bugs (intake writer + feature-agent parser) — currently queued behind two active PyAutoBrain tasks.
• Phase 1 — pre_build audit (draft/research/pyautobuild/pre_build_git_add_failure_audit.md): enumerate the git-add failure class, answer "what is pre_build's local commit for", decide the glob bug's fate (fix vs delete).
• Phase 2 — Heart evidence-chain audit (draft/research/pyautoheart/readiness_evidence_chain_audit.md): per-leg evidence map + what-is-GREEN-worth judgment + nightly-grant recommendation. After Phases 0–1.
• Phase 3 — env-profile/validation-gate redesign (draft/research/workspaces/env_profile_and_validation_gate_redesign.md): parallelisable with 1–2; reconcile with Phase 2 before finalising.
• Phase 4 — version-model fork (draft/bug/pyautobuild/release_version_sync_back_to_main.md): decide commit-backs vs mains-authoritative floors with Phases 1–3 evidence; evidence currently leans floors + fixing the exact-pin consumers.
• Phase 5 — agent failure-mode mitigations (draft/research/pyautobrain/agent_failure_modes_structural_mitigations.md): last, so every earlier phase's errors enlarge the catalogue first.

Phases are issued one at a time as each predecessor nears shipping; the three research briefs are Model: Fable, supervised, design-first.

Detailed implementation plan

Affected Repositories

• PyAutoBuild (primary / centre of gravity)
• PyAutoHeart, PyAutoBrain, PyAutoMind, workspaces (per phase)

Branch Survey (2026-07-16)

Repository	Current Branch	Dirty?
./PyAutoBuild	main	clean
./PyAutoHeart	main	clean
./PyAutoBrain	main	clean (claimed ×2: workspace-agent, wake-up-skill-rename)

Branching: the epic itself carries no branch; each phase gets its own task branch (feature/<phase-task>).

Cross-cutting constraints

• Never dispatch release.yml as a test (rehearsal defaults to false = real PyPI release).
• Never reintroduce git add -f for dataset/ (bug: release pipeline force-commits simulated workspace datasets (git add -f) #126/fix: honor .gitignore for dataset/ staging in pre_build (drop -f) [#126 leg 1] #150).
• Do not copy DISABLE_JAX reasoning across af/ag/al surfaces (use_jax defaults differ).
• Prefer deleting the trap to hardening or documenting it; every brief carries a "Trust nothing here" section — re-measure inherited numbers.

Key Files

• PyAutoMind/active/build_chain_umbrella.md — the umbrella prompt (full decomposition)
• PyAutoBuild/pre_build.sh, .github/workflows/release.yml — Phases 1, 4
• PyAutoHeart/heart/checks/test_run.py, heart/readiness.py — Phases 0, 2
• *_workspace_test/config/build/env_vars*.yaml, PyAutoBuild/autobuild/env_config.py — Phase 3

Original Prompt

Click to expand starting prompt

Build-chain umbrella: one coordinated fix across five filed briefs

Type: bug
Target: PyAutoBuild
Repos:

• PyAutoBrain
• PyAutoBuild
• PyAutoHeart
• PyAutoMind
• workspaces
  Difficulty: too-large
  Autonomy: supervised
  Priority: high
  Status: formalised

PyAutoBuild is the centre of gravity. This umbrella combines five already-filed
PyAutoMind draft briefs (plus their filed satellites) that all touch the
build/release/validation chain, so the work is sequenced as a single campaign
instead of five disconnected drafts. The umbrella's job is to sequence and
unify; each component brief remains the authoritative statement of its own
problem — read it in full before starting its phase.

Component briefs

1. draft/research/pyautobuild/pre_build_git_add_failure_audit.md
   (research, medium, Model: Fable) — pre_build.sh has exactly three
   failure-tolerant sites (:55, :79, :88) and all three are implicated: :79 was
   bug 1 (fatal under set -e, fixed PyAutoBuild#154); :88 is bug 2 (unmatched
   glob → git rejects the whole pathspec list, exit 128 swallowed by || true,
   nothing staged — filed, unfixed); :55 seds a README.rst no workspace has, so
   its || true is permanently load-bearing and a real sed failure is
   indistinguishable from the expected one. Three sites, three hits — the idiom
   is the suspect. Leads with the design question: release.yml regenerates and
   commits the same artifacts on the runner, so what is pre_build's local
   commit for? Redundant (delete it — the Skip EmceePlotter / NautilusPlotter under PYAUTO_TEST_MODE=2 #47 shape) or load-bearing in an
   unexercised case? Find out; do not assume the tidier answer. Known
   unverified claim inside: whether the README bump edits but goes uncommitted.

2. draft/bug/pyautobuild/release_version_sync_back_to_main.md (bug,
   medium) — a release leaves __version__ stamps and workspace pins frozen on
   main (by design since PyAutoBuild#120's floor model) while Colab URLs move,
   producing recurring "drift" alarms and two hand-done 15-repo bumps. The
   2026-07-15 findings reframe it as an undecided fork: (a) re-add commit-backs
   to main (rejected once — they caused stale CI storms and an org-wide cron
   pause) vs (b) keep mains authoritative and fix the three consumers that
   still read a floor as an exact pin (assistant --check-version, Heart
   version_skew, workspaces without floors). Evidence leans (b). The genuine
   bug either way: after 2026.7.9.1 the floors named the yanked
   2026.7.6.649 — a floor must always name an installable version.

3. draft/research/pyautoheart/readiness_evidence_chain_audit.md
   (research, large, Model: Fable) — four measured integrity failures in the
   evidence behind Heart's release verdict, found in one half-day: (1) the
   install-verification leg discarded every PASS for months (fixed, PR#77);
   (2) Heart's own test suite clobbers live ~/.pyauto-heart state (filed, see
   satellites); (3) the test_run leg's history is incomparable across runs —
   the smoke surface silently doubled and 30 failures trace to missing
   simulated datasets, possibly downstream of PyAutoBuild#150 dropping
   git add -f dataset/ (unestablished — settle it); (4) the Brain feature
   agent's path parser misroutes post-lifecycle-split prompts (now filed, see
   satellites). Core deliverables: a per-leg evidence map (artifact, writer,
   has the writer ever run, is the leg satisfiable at all) and a judgment on
   what the GREEN gate is worth today — whether the nightly's standing grant
   should stand, narrow, or pause. Its own instruction: read the pre_build
   audit first; findings 3 and 5 intersect it.

4. draft/research/workspaces/env_profile_and_validation_gate_redesign.md
   (research, large, Model: Fable) — redesign the env-profile + validation-gate
   system after one missing YAML line held nightly-release red for five nights.
   Eight observed failure modes (silent override, vacuous tests,
   hand-maintained lists, release config untestable at PR time, inconsistent
   siblings, silent pattern over-match, duplicated resolvers, layered duplicate
   env reads); already-tried-and-rejected list (smoke_tests.txt promotion,
   static guard, loud warning — killed by adversarial review; writing it down —
   refuted by a note that did not fire one day later). Target: a verified map,
   a design that makes the failure modes structurally hard, an incremental
   migration path, and an answer to the open question: is mode=release
   validating the NumPy path for most of the ag/al surface intended? Guard: do
   NOT copy autofit_workspace_test's DISABLE_JAX reasoning to ag/al, where
   use_jax defaults True and the var is genuinely load-bearing.

5. draft/research/pyautobrain/agent_failure_modes_structural_mitigations.md
   (research, large, Model: Fable) — the companion asking how to stop the
   agent producing these bugs: fourteen itemised errors from one session in
   five shapes (verification that confirms the assumption ×5, stored belief
   over repo ×2, recall instead of enumeration ×2, unvalidated tools ×3,
   shared-state/convention ×4). Load-bearing findings: a note naming the exact
   trap did not fire one day later (documentation is refuted as a primary
   mechanism); repos.yaml held every hand-listing miss and was never reached
   for. Hypothesis to attack: mechanisms that REFUSED an action caught ~100%,
   mechanisms that INFORMED caught 0%. Deliverables: validated taxonomy, ranked
   structural mitigations demonstrated against the catalogue, honest attack on
   the memory system, an estimate of the unnoticed-error set.

Filed satellites (in scope, referenced by the briefs)

• draft/bug/pyautoheart/test_suite_clobbers_live_heart_state.md (small,
  HIGH) — brief 3's finding 2, filed separately. Also a precondition:
  brief 3 forbids running Heart's suite against a live state dir until fixed.
• draft/bug/pyautobuild/root_level_git_add_stages_nothing_on_unmatched_glob.md
  (small) — brief 1's bug 2. Hold the fix: brief 1's design answer may
  delete the line instead of hardening it; fixing it first pre-empts the audit.
• draft/bug/pyautobrain/feature_agent_path_parser_predates_lifecycle_split.md
  (small, safe) — brief 3's finding 4, filed 2026-07-16 while decomposing this
  umbrella. Sibling: draft/bug/pyautobrain/intake_writes_legacy_layout.md
  (the writer side of the same pre-feat(python-version): support Python 3.9-3.14, first-class 3.12/3.13 #71 assumption).
• draft/feature/workspaces/minimum_library_version_adoption.md (small, HIGH)
  and draft/feature/pyautoheart/version_skew_floor_rework.md (small) — the
  two floor-model consumers brief 2's fork (b) would execute; the version_skew
  rework depends on the adoption landing first.

Task decomposition (worked out 2026-07-16)

Phase 0 — small unblockers (independent, land first, in any order):

• Fix test_suite_clobbers_live_heart_state (design-first per its brief:
  prefer run() returning the summary and only the CLI persisting; sweep
  heart/checks/ for siblings; decide deliberately on the autouse fixture).
  Unblocks Phase 2's method safely.
• Fix intake_writes_legacy_layout + feature_agent_path_parser_... (both
  small/safe Brain fixes; the second is the reader side of the first).

Phase 1 — pre_build audit (brief 1). First research phase: brief 3 says
read it first, and its design answer decides both the fate of the glob bug
(fix vs delete the line) and the #150→missing-datasets question brief 3's
finding 3 hangs on. Output: the failure-class enumeration, the
redundant-vs-load-bearing verdict, a target design (possibly a smaller file).

Phase 2 — Heart evidence-chain audit (brief 3). After Phase 1 (consumes
its #150 verdict) and Phase 0 (safe suite runs). Output: per-leg evidence map,
the what-is-GREEN-worth judgment, and the human decision on the nightly's
standing grant.

Phase 3 — env-profile/validation-gate redesign (brief 4). Parallelisable
with Phases 1–2 (different surface), but reconcile before finalising: its
answer to the ag/al mode=release NumPy question feeds brief 3's "what does the
verdict validate", and both feed the same release chain. Output: design doc +
incremental migration plan.

Phase 4 — version-model fork (brief 2). Decide fork (a)/(b) with Phases
1–3 evidence in hand; current evidence leans (b): land
minimum_library_version_adoption, then version_skew_floor_rework, rework
the assistant --check-version to be source-checkout-aware (or drop equality
for the API-surface hash), enforce "a floor names an installable version"
whichever fork wins, then rewrite or shelve the sync-back prompt with the
reasoning recorded. Never re-add stamp commit-backs without confronting the
#120 history (CI storms, cron pause).

Phase 5 — agent failure-mode mitigations (brief 5). Last, deliberately:
every earlier phase's "Trust nothing here" section, plus any new errors made
during Phases 0–4, are additional catalogue data; the refusals-vs-reminders
hypothesis should be tested against the enlarged set. Its mitigations then
protect all future campaigns.

Cross-cutting constraints (from the briefs; apply to every phase):

• Research briefs are Model: Fable, supervised, design-first — plan approval
  before implementation; judgment, not merged fixes, is the deliverable.
• Never dispatch release.yml as a test — rehearsal defaults to false = a
  real PyPI release. workspace-validation.yml mode=smoke is safe.
• Never reintroduce git add -f for dataset/ (bug: release pipeline force-commits simulated workspace datasets (git add -f) #126/fix: honor .gitignore for dataset/ staging in pre_build (drop -f) [#126 leg 1] #150).
• Do not copy DISABLE_JAX reasoning across af/ag/al surfaces (use_jax defaults
  differ).
• Prefer deleting the trap to hardening or documenting it.
• Each brief carries a "Trust nothing here" section: treat every inherited
  number as re-measurable and re-measure it.

[Jammy2211]

Phase 0a at PR-open: PyAutoLabs/PyAutoHeart#79 (fixes PyAutoLabs/PyAutoHeart#78 — test-suite state clobber + the version_skew sibling found by the empirical sweep). Unblocks Phase 2's method (Heart's suite can now run safely against live state).

Phase 0b (Brain lifecycle-path fixes) remains queued behind the two active PyAutoBrain claims. Next up once 0a merges: Phase 1, the pre_build audit (draft/research/pyautobuild/pre_build_git_add_failure_audit.md).

[Jammy2211]

Phase 0a MERGED (PyAutoLabs/PyAutoHeart#79; #78 closed). Heart's suite is now safe against live state — Phase 2's method is unblocked.

Phase 0b update: intake_writes_legacy_layout was fixed independently by a parallel session (record in Mind complete/2026/07/), so 0b reduces to the feature-agent parser fix — still queued behind the PyAutoBrain claims.

Phase 1 starting: the pre_build git-add failure-class audit (draft/research/pyautobuild/pre_build_git_add_failure_audit.md).

[Jammy2211]

Phase 1 COMPLETE — #157 (audit) + #158 (step 1) + #159 (step 2) all merged; step 3 parked in Mind planned.md on the dataset/config-sweep human decision (audit doc §6). Phase 2 open: PyAutoLabs/PyAutoHeart#83 (readiness evidence-chain audit), task heart-evidence-audit.

[Jammy2211]

Phase 1 step 3 at PR-open (PyAutoBuild#160 — sweep vestige DELETED per human decision) and Phase 2 audit at PR-open (PyAutoLabs/PyAutoHeart#84). Phase 2's finding-3 remediation filed as a new satellite (draft/bug/workspaces/howto_validation_needs_simulator_stage.md). Awaiting human merges of #160 and Heart#84; Phase 2 fix PRs (test_run wiring, verify_install exercise, surface recording, version_skew floor rework → Phase 4) follow behind plan approval.

[Jammy2211]

Phase 0b at PR-open (PyAutoLabs/PyAutoBrain#128 — active/ header fallback; the draft/ half was independently fixed in Brain 9f27c5f this morning) and Phase 2 fix 1 at PR-open (PyAutoLabs/PyAutoHeart#85 — test_run caller-decided fetch_cloud + local/cloud agreement). Phases 0-2 are now fully shipped or awaiting merge; next up: Phase 3 (env-profile/validation-gate redesign brief) and Phase 4 (version-model fork, now holding the version_skew rework + README-pin ownership + floors-name-installable-versions invariant).

[Jammy2211]

Phase 3 design at PR-open (#162) on the human-decided invariant, and Phase 4 fork DECIDED (b) — mains stay authoritative; the decision + 4-task execution programme are recorded in draft/bug/pyautobuild/release_version_sync_back_to_main.md (floors adoption → version_skew rework incl. the installable-floors invariant → assistant --check-version fix → README-pin ownership; serialised on the autolens_workspace claim). Remaining after these: Phase 5 (agent failure-mode mitigations — its catalogue has grown considerably this campaign) and the Phase 2 satellites (verify_install exercise run, surface recording, HowTo* simulator stage).

[Jammy2211]

Phase 5 at PR-open (PyAutoLabs/PyAutoBrain#131 — the capstone doc: validated + enlarged catalogue, sharpened refusals hypothesis, six ranked mitigations as future tasks). Phase 3 step 1 MERGED (#163 — validator live; found 3 dead patterns + 15 vacuous-JAX scripts on first contact). verify_install evidence run in progress (Phase 2 finding C — first post-fix exercise of the leg). Campaign status: no unstarted phases remain; all remaining work is enumerated mechanism/migration tasks.