feat(sc-git-worktree, sc-commit-push-pr): stack-aware worktree factory and commit/PR flows (0.13.0) - #107
Open
randlee wants to merge 13 commits into
Open
feat(sc-git-worktree, sc-commit-push-pr): stack-aware worktree factory and commit/PR flows (0.13.0)#107randlee wants to merge 13 commits into
randlee wants to merge 13 commits into
Conversation
randlee
force-pushed
the
feature/sc-git-worktree-gh-stack
branch
from
August 30, 2026 23:35
30a2df4 to
3f6129b
Compare
…stack Surgical additions to the tested package, opus-reviewed twice (review -> fixes -> empirical verification incl. reproducing the destroy scenario pre/post-fix): - references/gh-stack-support.md (gated): defers ALL stacked-PR authority to the managing-gh-stacks skill (sc-gh-stack) when installed; strongly recommends installing it otherwise; 5-row non-interactive survival-kit table as fallback, banner-marked as not a substitute for the skill - SKILL.md: one description clause + one gated pointer section. Two independent gates: destructive safety is UNCONDITIONAL (stack tracking is repository state, not local toolchain); the extension gate applies only to stack operations - Destructive-safety guard (the one unavoidable code change): batch cleanup skips any worktree carrying gh-stack tracking (.git/worktrees/<wt>/gh-stack), surfaced in gh_stack_skipped — verified: pre-fix batch cleanup destroyed a stack worktree whose bottom layer looked merged (tracking lost, open PRs above orphaned); post-fix it is preserved. scan reports gh_stack_tracked per worktree + summary count; single-branch cleanup/abort surface it without blocking. check_gh_stack_tracked fails closed incl. OSError on prunable worktrees (regression caught in verification) - Pre-existing doc fixes: abort agent protected-branch clause matched to implementation; TRACKING.MISSING casing - 12 new tests (132 -> 144), all validators green Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The invariant: a worktree must be stacked iff it NEEDS to be — the base branch decides. Basing on trunk/protected/merged = independent work, flat create proceeds byte-identically (the many existing /sc-git-worktree prompts are unaffected). Basing on an unmerged branch = the work depends on unmerged work = a stack layer by definition: worktree_create.py refuses flat creation (CREATE.NEEDS_STACK, before any mutation) and routes — into the tracked stack's own worktree when the base carries gh-stack tracking (layers share one worktree), else a new stack/<base-slug> worktree + gh stack init. Explicit flat:true overrides. Above this hard floor the docs add the activity layer: stack-oriented work creates stacked even off trunk (orchestrator judgment, not script-enforced). Opus-verified empirically (full matrix incl. no-mutation-on-refusal and routing), with its blocker finding fixed: the guard fails OPEN when protected-branch config is absent, so config-free repos keep the historical flat-create path (regression test added). Also fixed in passing, verified safe: is_branch_merged missed git's '+ ' prefix for branches checked out in other worktrees — single-branch cleanup of a genuinely merged branch always refused before; batch cleanup proven byte-identical pre/post. suggested_worktree_path honors a custom worktree_base; cleanup agent doc notes single-branch mode is the only path that can remove a tracked stack worktree. 7 new tests (144 -> 151); validate-all 9/9 PASS. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…accreted guards) Designed rework per DESIGN.md 'Worktree factory decision model': create resolves every request to exactly one product — A (flat, legacy, byte-identical to pre-guard behavior, verified against git history), B (new stack: same path as flat + rerere + gh stack init, requested_base surfaced when policy overrides the base), C (layer: no new worktree; checkout + gh stack add inside the base's stack worktree). Precedence Intent > Dependency > Policy > default A, evaluated lazily under the positive-signal rule: a repo that is not stack-active (no always_stack, no gh-stack tracking anywhere) never evaluates any of it — legacy prompts are auto-upgraded only where stacks demonstrably exist, and branch-of-branch flat creates keep working in stack-naive repos. Mandatory collaborator prereq gate (gh + extension + sc-gh-stack skill) fires only in stack-active repos, before any mutation. Fallback settings parser strips inline comments; unknown boolean scalars coerce False; pyyaml declared in requires. Tests rebuilt from the decision matrix: test_factory_decision.py (24) replacing the two incremental guard suites; 167 total green; validate-all 9/9. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…ndency This package is the critical junction where stack-unaware pushes, PRs, and destination-merges corrupt stacks — per design it now depends on the gh-stack toolchain outright (manifest: gh CLI, gh-stack extension, sc-gh-stack package). New scripts/stack_guard.py (self-contained, no cross-package imports): unconditional prerequisite gate at every entry (PREFLIGHT.STACK_PREREQS_MISSING with exact installs — wired into preflight so agent start-hooks block before any commit), and state-based stack-layer detection (.git/worktrees/<wt>/gh-stack). On a layer: destination pull/merge is SKIPPED (stack linearity; syncing belongs to gh stack sync), push and PR creation refuse with STACK.USE_GH_STACK routing to the managing-gh-stacks skill by name, committed/pushed state reported accurately. Non-stack flows byte-compatible once prereqs pass; Azure DevOps flows untouched by construction. Agents/SKILL/README document the codes and the never-work-around rule. 26 new tests. Version 0.13.0 (synchronized marketplace release; registries regenerate at the release layer). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Both packages carry behavior changes in this stack layer and ship at the synchronized marketplace release version (matching PR #84's platform bump and sc-gh-stack's 0.13.0 below); registries regenerated via set-package-version.py. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…, JSONL-as-trace)
Additive codex path: gpt-*/luna/sol/terra models route through codex
exec --yolo --json; command_execution/agent_message/usage events map
into CollectedData so every existing expectation evaluator works
unchanged; {test_id}-codex-events.jsonl is the trace artifact (codex has
no hooks). CODEX_HOME provisioned inside the isolated HOME (auth via
copy_codex_auth); ZDOTDIR-owned dotfiles keep stub PATH control under
zsh -lc; AGENTS.md generated as codex's skill channel (Skill-tool
equivalent). Claude path byte-identical (regression-tested). 35 new
tests; harness suite 671 -> 706. Companion docs in sc-test-harness
docs/CODEX.md.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Six cases mapped to DESIGN.md decision-table rows, non-stack AND stacked: flat-create-stays-flat, branch-of-branch-naive (distribution guarantee), always-stack-product-b, dependent-base-joins-stack (product C, no new worktree), prereqs-missing-onboarding, cleanup-skips-tracked- stack. All scaffolds mechanically verified against the real scripts (two fixtures initially hit the positive-signal fail-open exactly as designed — fixed with explicit protected-branches config). Skill tool allowed in every case. Fixtures generated + FixtureLoader-verified. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Four cases: non-stack-flow-normal (control, byte-compat), stack-layer-refusal-routing (no push/PR workaround; routes to managing-gh-stacks by name), prereqs-gate-onboarding (exact install lines, no improvisation), no-destination-merge-on-layer (poison-file grader proven to discriminate by simulating the bad-agent merge). Mechanical finding: url.insteadOf can't fake a GitHub remote for provider_detect (get-url applies substitution) — bin/git passthrough wrapper intercepts only 'remote get-url origin'. Fixtures generated + FixtureLoader-verified. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
codex exec with a piped/inherited stdin waits to read it as an appended
<stdin> block ('Reading additional input from stdin...') and never
starts the turn — found empirically when backgrounded eval reruns
produced zero events until timeout. Same fix as the local runner.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…6 on worktree Cross-model prove-out iterations (full trail in site/reports/evals): - Runner unified across worktrees (each copy had half the fixes) + stdin=DEVNULL and partial-events capture on the claude path too; absolute workspace-bin PATH prepend (stubs win for script-spawned processes in other cwds — turned always-stack-product-b green) - git-wrapper scaffolds: canonicalize PATH entries before self-exclusion (relative ./bin WAS the wrapper -> infinite self-exec hang at claude startup) - Partial-events capture found two more hang causes: model-chosen background bash tasks stall headless -p sessions (prompts now forbid backgrounding), and transient 401 auth throttling under rapid batches - SKILL script-route rules from graded failures: default sibling layout unless told otherwise; sweeps use BATCH cleanup (single-branch removal of gh_stack_tracked:true requires explicit approval — stop and ask); relay refusal suggested_action verbatim, never soften into workaround offers. Agent-tool naming sweep (formerly Task); codex collaboration.spawn_agent delegation notes - Grader calibration: dependent-base checked-out grader removed (redundant + brittle; outcome pinned by no-new-worktree-dir + stack-add-ran + llm judge) Scorecard: sc-git-worktree luna 6/6; haiku oscillates on 3 multi-step flows across runs — recorded as a model-reliability finding (the skill content is proven by luna's consistency), not a skill defect. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…ity, residue scrub
The isolated-run debugging arc, each fix transcript-verified:
- pytest_plugin: implement setup.files (FileMapping) execution — loader
supported it, execution never wired, so file-mapped fixtures (incl.
all generated eval scaffolds) silently never arrived in the project
- environment: validate ANTHROPIC_API_KEY once per process and scrub
only provably-dead keys (a stale env key hijacks fresh-HOME sessions
with 401s; OAuth does not transfer into an isolated HOME by file copy
— a VALID key is the isolated-auth path); copy_claude_auth for
credential-file setups
- runner: collected_data.cwd set authoritatively to the project path
(transcript cwd does not survive HOME-override translation)
- generator: per-case residue scrub before scaffolds (harness project
persists across cases); escape-processing fix in the frontmatter
parser (authored regexes like STACK\\.X arrived double-escaped and
could never match — six graders across suites were inert)
- sc-commit-push-pr eval budgets raised (max_turns 40 / 450s) — the
pipeline suite legitimately needs more room
Companion fix in sc-test-harness (dbb69de): hook commands pinned to
${CLAUDE_PROJECT_DIR} — hooks run in the shell's cwd, so any agent cd
broke every relative log-hook.py invocation, and PreToolUse blocking
errors froze all tool execution. Harness suite 706 green.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…arness runs Full progression: sc-git-worktree haiku iterations + luna 6/6; sc-commit-push-pr local-runner iterations (environment-blocked era) and the first isolated harness runs, ending at pure behavioral results: haiku breaks script-first discipline (direct push on a stack layer, speculation instead of running the gate) where luna holds it — the measured model-reliability floor. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…ok tests; CI greens The unconditional gh-stack prerequisite gate (this branch's sc-commit-push-pr change) fires before every scenario in the pre-existing repo-level suites — tests/test_commit_push.py, test_create_pr.py, test_hooks.py now satisfy it via an autouse fixture (gate behavior itself is covered by the package's test_stack_awareness.py). tests/test_sc_prefix_validation.py pins sc-git-worktree at 0.13.0 (bumped ahead of the platform release at this stack layer; refactory's bump above moves _CURRENT_VERSION and reconciles). Eval scaffolds chmod +x (security-scan). Full repo suite green; validate-all ALL PASSED. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
randlee
force-pushed
the
feature/sc-git-worktree-gh-stack
branch
from
August 31, 2026 05:26
ff9a529 to
da45e55
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Middle layer of the stack:
develop <- #101 (sc-gh-stack) <- this <- #84 (refactory).sc-git-worktree — worktree factory decision model
Create resolves every request to exactly one product (see the package's DESIGN.md):
gh stack init(repo policygit.always_stack,stack_rootdefault develop).gh stack add(dependency: an unmerged base is by definition a stack layer).Precedence
Intent > Dependency > Policy > default A. Mandatory collaborator prerequisites in stack-active repos (CREATE.STACK_PREREQS_MISSINGlists installs). Destructive safety: batch cleanup skips gh-stack-tracked worktrees; scan/cleanup/abort reportgh_stack_tracked. Optional gh-stack interop reference defers all stack authority tomanaging-gh-stacks.sc-commit-push-pr — fully stack-aware, hard gh-stack dependency
This package is the critical junction where stack-unaware pushes/PRs/merges corrupt stacks:
STACK.USE_GH_STACKrouting tomanaging-gh-stacksby name, accurate committed/pushed reporting.Both packages version-matched at 0.13.0 (synchronized marketplace release; registries regenerated per layer). Tests: 167 (sc-git-worktree) + 26 (sc-commit-push-pr); validate-all 9/9. Opus-reviewed with empirical verification across four rounds.
🤖 Generated with Claude Code