Revert Claude Code substrate

[jwbron]

Summary

Abandons the Claude Code substrate initiative. Reverts the three commits that landed substrate code/docs on main:

• 802f77d (Spike: walking-skeleton Claude Code substrate for egg SDLC (#2623) #2715) — walking-skeleton spike
• 173f483 (docs: add claude-code substrate to index and structure docs [doc-updater] #2718) — substrate docs in index/structure docs
• 82c4ba4 ([#2717] Substrate-swap follow-up: integration branch #2731) — substrate-swap rollout integration (slices 1–2, [slice-1] Roll out Claude Code substrate to remaining roles + plan/... #2724/[slice-2] Roll out Claude Code substrate to remaining roles + plan/... #2726)

Removes orchestrator/substrate/, plugins/egg-sdlc/, the claude-code-substrate ADR, associated unit/integration tests, and the #2717/#2623 .egg-state pipeline artifacts. 73 files, −65,257 lines.

The four substrate PRs (#2715, #2724, #2726, #2731) were already merged, so they cannot be closed — this revert is how the work is removed. Related tracking issues (#2717, #2729, #2733, #2734, #2735, #2736) are being closed alongside this PR.

Test plan

• make test — 17887 passed, 41 skipped (changeset-aware suite green; no main code depends on the removed substrate modules)
• No remaining references to orchestrator.substrate / plugins/egg-sdlc / claude-code-substrate in code or docs
• Reverts applied with no merge conflicts

[egg-reviewer]

No agent-mode design concerns. This is a pure revert (65K lines removed, ~2 modified in non-deleted files) of the Claude Code substrate initiative; it does not introduce any new prompts, agent-dispatch patterns, pre-fetching, post-processing pipelines, structured-output-for-humans surfaces, hardcoded model IDs, or direct API calls. Whether to abandon the substrate strategy is an operator decision, not an agent-mode design question.

— Authored by egg

[egg-reviewer]

Review: clean wholesale revert, approve

Verified this is a faithful three-commit revert (802f77d, 173f483, 82c4ba4) with no collateral changes and no dangling references.

What I checked

Revert completeness — searched the whole tree (excluding .egg-state/) for the removed surfaces (EGG_SUBSTRATE, SubstrateBundle, select_substrate, ClaudeCodeSpawner, K3sSpawnerAdapter, run_pipeline_in_process, InProcessMessageBus, PreToolUseHookPolicy, LocalWorktreeManager, orchestrator.substrate, plugins/egg-sdlc, claude-code-substrate, the literal word substrate). Zero hits in .py, .md, .toml, .yaml, .yml, .json, Makefile, .github/, or scripts/.

Collateral check on the three reverted commits:

• 802f77d — only substrate code/tests/docs + .egg-state/ artifacts. The conftest.py touch was the substrate fixture plus a small "fix-along" on two except clauses (see note below).
• 173f483 — only docs/development/STRUCTURE.md + docs/index.md, both substrate-only.
• 82c4ba4 — only substrate code/tests/docs + .egg-state/ artifacts.

No unrelated work is being lost.

Surgical modified files (the ones that aren't pure deletions):

• orchestrator/concurrent_executor.py — _spawn_agent is restored to its pre-substrate form (legacy self.spawn_fn(...) path only); _spawn_agent_via_substrate / _teardown_role_worktree are removed cleanly; no orphan imports — all of AgentExecution, AgentExecutionStatus, Path, UTC, datetime are still used elsewhere in the file.
• integration_tests/regression/conftest.py — substrate fixture + _in_sandbox_agent_context helper removed; __all__ updated; os still used elsewhere in the file (no orphan import).
• The three docs files: only the substrate bullet/section is removed; rest of the file is untouched.

Orphan tests: none. None of the remaining tests reference the substrate fixture, from orchestrator.substrate, or any of the removed test helper names.

Pre-existing egg-sdlc references in docs and pyproject.toml are to the separate bin/egg-sdlc CLI script and to the plugins/refine-plan/ plugin, not to the removed plugins/egg-sdlc/ plugin. Not orphans.

Non-blocking note

integration_tests/regression/conftest.py:283,293 — the revert reintroduces the un-parenthesized form:

except OSError, subprocess.TimeoutExpired:
    return None
...
except ValueError, UnicodeDecodeError:
    return None

This is valid Python 3.14 syntax per PEP 758, which the project's requires-python = ">=3.14" floor allows (verified with py_compile on 3.14.5). It is not a syntax error and the file imports cleanly.

That said, this form will read as a Python 2 syntax error to almost every reviewer who sees it (the original substrate PR added # fmt: skip parens specifically to avoid that confusion). Since the only purpose of the revert here is to match pre-substrate state, I'd suggest leaving those two paren-wraps in place — it's a 2-line readability win that has nothing to do with the substrate work and would otherwise be the kind of nit a future reviewer flags. Strictly optional.

Verdict

LGTM. The revert is mechanically clean, eliminates all live references to the removed surfaces, and preserves the pre-substrate k3s path verbatim in concurrent_executor.py. The make test result the PR description cites (17887 passed / 41 skipped) is consistent with what I'd expect given that nothing on main outside the reverted commits referenced these modules.

— Authored by egg

[james-in-a-box]

Review response

Both bot reviews approved; only one actionable item was raised.

Disposition

• agent-mode-design review — approve, no concerns raised, nothing to address.
• review — approve-with-suggestions. Non-blocking note on the un-parenthesized except clauses at integration_tests/regression/conftest.py:283,293:
  • fixed-in-PR (commit 3c14f60) — re-added the (OSError, subprocess.TimeoutExpired) and (ValueError, UnicodeDecodeError) paren-wraps. The "non-blocking" framing doesn't count as an explicit defer request per the feedback contract, and the fix is a 2-line change with no risk.

— Authored by egg

[james-in-a-box]

Autofix tracking

{"Lint/Python": 1}

[egg-reviewer]

Re-review at 1f2dea7: no agent-mode design concerns.

Delta since the previous review (a38bfbd) is two commits (3c14f60, 1f2dea7) that only touch integration_tests/regression/conftest.py and net to zero tree changes — the auto-formatter commit reverted the paren-wrap re-add. Nothing in the delta touches agent prompts, dispatch, pre-fetching, structured-output surfaces, model IDs, or direct API call paths.

— Authored by egg

[egg-reviewer]

Re-review at 1f2dea7: approve, no new issues

Delta since previous review

Two commits added since a38bfbd:

1. 3c14f60 — "Parenthesize except tuples in lifecycle_secret" (egg-reviewer[bot])

   • Re-added the paren-wraps in integration_tests/regression/conftest.py:283,293 per the non-blocking note from my previous review.

2. 1f2dea7 — "Fix checks: apply automated formatting fixes" (egg)

   • The exact same two lines, reverted: ruff (configured with target-version = "py314" + UP rules) prefers the PEP 758 bare form on 3.14+ and unwrapped the parens.

git diff a38bfbd..1f2dea7 is empty — the two commits cancel out. The tree is byte-identical to what I approved before.

Verification

• git diff --name-only prev-review pr-2742 → no files changed.
• Verified the bare except OSError, subprocess.TimeoutExpired: form is valid syntax on Python 3.14.5 (PEP 758), and the project pins requires-python = ">=3.14", so this runs.

On the unresolved non-blocking suggestion

My prior "leave the parens for readability" note can't actually be addressed without lint-config changes — ruff at py314 will keep removing them. If the team wants to enforce parens for human readability, the fix lives in pyproject.toml ([tool.ruff.lint.per-file-ignores] for whichever UP rule unwraps them, or a project-wide ignore), not in the source files. That's well outside the scope of a revert PR.

I am not asking for any change here. Strictly non-blocking, and tracking it separately from this revert is the right call.

Verdict

LGTM. Same approval as before. Net-zero delta means no regression risk introduced by the re-push.

— Authored by egg

[james-in-a-box]

egg review completed. View run logs

7 previous review(s) hidden.