Promote develop to main (200 PRs, 160+ issues, 179 fixes, 480 features, 27 refactors, 22 infra)#2213
Conversation
…wn (#1961) ## Summary Add a `Model` column to the per-step token summary table and a new per-model aggregate breakdown table. The model identity is sourced from the `model_breakdown` dict already parsed by `extract_token_usage()` in `_session_model.py`. The change threads model identity through three paths: (1) in-memory accumulation via `TokenEntry.model`, (2) on-disk persistence via a new `model_identifier` field in `token_usage.json`, and (3) formatting in `TelemetryFormatter`, the stdlib-only hook, and the compact PostToolUse formatter. No cost estimation is included — the acceptance criteria require token counts only, and no pricing infrastructure exists. Closes #1906 ## Implementation Plan Plan file: `/home/talon/projects/autoskillit-runs/impl-20260505-135150-264911/.autoskillit/temp/make-plan/token_summary_model_column_plan_2026-05-05_135500.md` 🤖 Generated with [Claude Code](https://claude.com/claude-code) via AutoSkillit <!-- autoskillit:pipeline-signature steps=prepare_pr,run_arch_lenses,compose_pr,annotate_pr_diff,review_pr --> --------- Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…ll-Count Tests (#1967) ## Summary Remove redundant and brittle test code across three test modules: parametrize the three identical kitchen_rules rejection tests, delete the duplicate session-type warning test, and convert exact skill-count assertions to lower-bound checks. No production code changes. Closes #1886 ## Implementation Plan Plan file: `/home/talon/projects/autoskillit-runs/impl-20260505-180602-586205/.autoskillit/temp/make-plan/deduplicate_session_type_kitchen_rules_skill_count_tests_plan_2026-05-05_180602.md` 🤖 Generated with [Claude Code](https://claude.com/claude-code) via AutoSkillit <!-- autoskillit:pipeline-signature steps=prepare_pr,run_arch_lenses,compose_pr,annotate_pr_diff,review_pr --> --------- Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
…1969) ## Summary The `resolve-failures` SKILL.md has an ambiguous verdict decision flow that allows an LLM executor to emit `ci_only_failure` even after successfully applying a fix. The fix restructures the Step 2d decision tree to make the override rule explicit: **any time a code change is committed and tests pass, the verdict is `real_fix`, regardless of `failure_subtype`**. The Step 2d table is clarified to apply ONLY to the "no fix applied" path, and a post-fix-loop verdict override is added to prevent re-evaluation through the wrong decision path. ## Requirements - REQ-RF-001: When `resolve-failures` applies a code change AND the subsequent CI run passes, the verdict MUST be `real_fix`, not `ci_only_failure` - REQ-RF-002: `ci_only_failure` should only be emitted when no fix was applied or when the applied fix did not resolve the CI failure - REQ-RF-003: The fix must not break the existing `ci_only_failure` path for genuinely unfixable CI failures Closes #1954 ## Implementation Plan Plan file: `/home/talon/projects/autoskillit-runs/impl-20260505-180603-520539/.autoskillit/temp/make-plan/resolve_failures_ci_only_failure_verdict_fix_plan_2026-05-05_181000.md` 🤖 Generated with [Claude Code](https://claude.com/claude-code) via AutoSkillit <!-- autoskillit:pipeline-signature steps=prepare_pr,run_arch_lenses,compose_pr,annotate_pr_diff,review_pr --> --------- Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
…itHub API Usage) (#1970) ## Summary Add a `local_review_rounds` configuration and plumbing so that the existing review loop steps (`annotate_pr_diff`, `review_pr`, `resolve_review`, `check_review_loop`) receive a `review_mode` context value (`"local"` or `"github"`) computed per-iteration. Part A covers the config dataclass, defaults, ingredient bridge, callable modification, recipe YAML wiring for all three looping recipes, and tests for all of the above. Part B will cover the skill SKILL.md behavioral changes (how `review-pr` and `resolve-review` branch on `mode=local` vs `mode=github`) — implement as a separate task. Closes #1945 ## Implementation Plan Plan file: `/home/talon/projects/autoskillit-runs/impl-20260505-165914-350121/.autoskillit/temp/make-plan/local_review_rounds_plan_2026-05-05_170500_part_a.md` 🤖 Generated with [Claude Code](https://claude.com/claude-code) via AutoSkillit <!-- autoskillit:pipeline-signature steps=prepare_pr,run_arch_lenses,compose_pr,annotate_pr_diff,review_pr --> --------- Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
…ng (#1971) ## Summary Update three SKILL.md files (`compose-pr`, `prepare-pr`, `diagnose-ci`) to use directive description language and fix a step numbering gap in `diagnose-ci`. Then uncomment four step overrides in `.autoskillit/config.yaml` to route `retry_worktree`, `compose_pr`, `diagnose_ci`, and `prepare_pr` to the MiniMax M2.7-highspeed profile. ## Requirements ### REQ-1: Update 3 SKILL.md descriptions to directive language PR #1937 updated 6 SKILL.md files to use directive language (per the Seleznov 650-trial study: 94-100% activation vs 37-77% for passive descriptions). Three of the four new MiniMax target skills still use passive descriptions. **Files to update:** | File | Current description style | Required change | |------|--------------------------|-----------------| | `src/autoskillit/skills_extended/compose-pr/SKILL.md` | Passive prose: "Reads the PR prep file and validated arch-lens diagrams..." | Directive: "PR composition executor. ALWAYS invoke this skill when instructed to compose a PR. Do not read prep files or create PRs directly — use this skill first to load the composition workflow." | | `src/autoskillit/skills_extended/prepare-pr/SKILL.md` | Passive prose: "Reads plan(s), runs git diff, classifies changed files..." | Directive: "PR preparation executor. ALWAYS invoke this skill when instructed to prepare PR metadata. Do not read plans or classify files directly — use this skill first to load the preparation workflow." | | `src/autoskillit/skills_extended/diagnose-ci/SKILL.md` | No `description:` field at all | Add directive: "CI diagnosis executor. ALWAYS invoke this skill when instructed to diagnose CI failures. Do not fetch CI logs directly — use this skill first to load the diagnosis workflow." | `retry-worktree` already has directive language (updated in PR #1937). **Why this matters:** The hook-based skill load guard (PR #1937) enforces Skill tool loading regardless of description language. But directive descriptions improve voluntary model compliance — belt-and-suspenders. MiniMax's "thoughtful disobedience" pattern means every compliance signal helps. ### REQ-2: Fix diagnose-ci step numbering gap The `diagnose-ci` SKILL.md workflow skips from Step 1 to Step 3 — there is no Step 2. This is likely accidental (no logical reason for the gap). MiniMax may attempt to invent a Step 2 from training priors, causing unintended tool calls. **File:** `src/autoskillit/skills_extended/diagnose-ci/SKILL.md` **Current numbering:** Step 1, Step 3, Step 4, Step 5, Step 5a, Step 6, Step 7 (no Step 2). **Fix:** Renumber to sequential: Step 1, Step 2, Step 3, Step 4, Step 4a, Step 5, Step 6. | Old | New | |-----|-----| | Step 3: Fetch Failure Summary | Step 2: Fetch Failure Summary | | Step 4: Fetch Per-Job Logs | Step 3: Fetch Per-Job Logs | | Step 5: Classify Failure | Step 4: Classify Failure | | Step 5a: Subtype Classification | Step 4a: Subtype Classification | | Step 6: Write Diagnosis Report | Step 5: Write Diagnosis Report | | Step 7: Emit Output Tokens | Step 6: Emit Output Tokens | **Bonus fix:** Line 80 says "proceed to Step 5 (write minimal diagnosis)" — but old Step 5 is Classify Failure, not Write Diagnosis. After renumbering, "proceed to Step 5" correctly resolves to Write Diagnosis Report for the first time. This fixes a latent cross-reference bug. ### REQ-3: Verify step override activation After REQ-1 and REQ-2 are complete, uncomment the 4 step overrides in `.autoskillit/.secrets.yaml` and run one implementation pipeline to verify: 1. Each step receives the FIRST ACTION directive (check session JSONL for "FIRST ACTION" in the prompt) 2. Each step calls the Skill tool as its first action (check JSONL for `"name":"Skill"` as the first tool call) 3. Each step completes successfully with structured output tokens emitted correctly Closes #1966 ## Implementation Plan Plan file: `/home/talon/projects/autoskillit-runs/impl-20260505-185537-800888/.autoskillit/temp/make-plan/prepare_4_run_skill_steps_for_minimax_m27_routing_plan_2026-05-05_190500.md` 🤖 Generated with [Claude Code](https://claude.com/claude-code) via AutoSkillit <!-- autoskillit:pipeline-signature steps=prepare_pr,run_arch_lenses,compose_pr,annotate_pr_diff,review_pr --> --------- Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
…#1972) ## Summary The `run_python` MCP tool dispatches callables through `_import_and_call` using a `dict[str, object]` parameter — a type-erasure boundary. Direct MCP tools get Pydantic lax-mode validation via FastMCP (which coerces `str→int` automatically), but `run_python` callables receive raw unvalidated values. The dispatcher already calls `inspect.signature(func)` but only uses it for `None→default` coercion (PR #1602), never reading `param.annotation`. This creates an asymmetry where `str`-typed callable parameters receiving JSON integers crash at subprocess boundaries, f-string operations, or Path construction. The architectural fix: extend `_import_and_call`'s existing coercion loop to read `param.annotation` and coerce primitive scalar types, closing the type-safety gap between the two dispatch paths. Defense-in-depth: add `str()` guards at each subprocess call site. Structural tests: a parametrized test matrix that exercises every callable × every param type combination through `_import_and_call`. Closes #1965 ## Implementation Plan Plan file: `/home/talon/projects/autoskillit-runs/remediation-20260505-183247-696032/.autoskillit/temp/rectify/rectify_run_python_type_coercion_2026-05-05_183800.md` 🤖 Generated with [Claude Code](https://claude.com/claude-code) via AutoSkillit <!-- autoskillit:pipeline-signature steps=prepare_pr,run_arch_lenses,compose_pr,annotate_pr_diff,review_pr --> --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
… State (#1968) ## Summary When a fleet campaign dispatch fails, the `FAILURE` status is terminal — `_ALLOWED_TRANSITIONS[FAILURE]` is `frozenset()` with no outgoing transitions. This blocks explicit user retry (`--resume`) because both the `has_failed_dispatch()` halt guard in `dispatch_food_truck` and Phase 2 of `resume_campaign_from_state` unconditionally reject campaigns with any FAILURE record. The fix adds a `FAILURE → PENDING` transition, a `reset_failed_dispatch()` function, and modifies the two halt check sites to distinguish between **automatic continuation** (should still halt) and **explicit user retry** (should reset the failed dispatch and re-execute). ## Requirements - REQ-RETRY-001: A failed dispatch MUST be retryable without manual state file edits - REQ-RETRY-002: The halt-on-failure guard MUST still prevent automatic continuation to subsequent dispatches after an unacknowledged failure - REQ-RETRY-003: Retry of a failed dispatch MUST reset its state and re-execute it from scratch (not resume) - REQ-RETRY-004: The retry mechanism MUST be safe under concurrent access (respect existing `_resume_lock` + `fcntl.LOCK_EX` pattern) Closes #1695 ## Implementation Plan Plan file: `/home/talon/projects/autoskillit-runs/impl-20260505-180604-269785/.autoskillit/temp/make-plan/fleet_campaign_retry_blocked_by_terminal_failure_state_plan_2026-05-05_181500.md` 🤖 Generated with [Claude Code](https://claude.com/claude-code) via AutoSkillit <!-- autoskillit:pipeline-signature steps=prepare_pr,run_arch_lenses,compose_pr,annotate_pr_diff,review_pr --> --------- Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
…1973) ## Summary Create `src/autoskillit/recipes/research-review.yaml` as a standalone sub-recipe containing the 22-step PR/review phase extracted from `research.yaml`. The recipe receives campaign-injected hidden ingredients (`worktree_path`, `research_dir`, `report_path`, `experiment_plan`, `experiment_results`, `experiment_type`, `scope_report`, `visualization_plan_path`), lifts all review steps verbatim-and-adapted, replaces the archival phase with dual terminal stops (`review_pr_complete` for PR mode, `review_local_complete` for local mode), and corrects routing targets to terminate within this sub-recipe rather than routing to archival or non-existent steps. Closes #1702 ## Implementation Plan Plan file: `/home/talon/projects/autoskillit-runs/impl-20260505-184855-274993/.autoskillit/temp/make-plan/p2_wp3_create_research_review_yaml_plan_2026-05-05_185500.md` 🤖 Generated with [Claude Code](https://claude.com/claude-code) via AutoSkillit <!-- autoskillit:pipeline-signature steps=prepare_pr,run_arch_lenses,compose_pr,annotate_pr_diff,review_pr --> --------- Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
…1974) ## Summary _INFRA_UNCONDITIONAL_FILES in tests/_test_filter.py contains 9 filenames, but 3 of them (test_hook_executability.py, test_hook_registration_coverage.py, test_hook_registry.py) live in tests/hooks/, not tests/infra/. The path construction loop at line 1271 resolves all 9 under tests/infra/, silently dropping the 3 hook tests from every tiered conservative filter run. The guard test in test_test_filter_tiered_always_run.py checks only basenames (p.name), masking the bug. This was introduced by commit 26c8059 (#1734) which moved the files without updating the constant. The fix splits the constant into two frozen sets with correct directory mappings, adds a second path construction loop, and strengthens all guard tests to assert parent directories. ## Implementation Plan Plan file: `/home/talon/projects/autoskillit-runs/impl-20260505-203325-376042/.autoskillit/temp/make-plan/fix_test_filter_hook_test_path_mismatch_plan_2026-05-05_203500.md` ## Changed Files - tests/_test_filter.py - tests/test_test_filter.py - tests/test_test_filter_coverage_map.py - tests/test_test_filter_tiered_always_run.py Closes #1875 🤖 Generated with [Claude Code](https://claude.com/claude-code) via AutoSkillit <!-- autoskillit:pipeline-signature steps=prepare_pr,run_arch_lenses,compose_pr,annotate_pr_diff,review_pr --> --------- Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
…1977) ## Summary Fix two validated audit findings (C4-1 and C4-3) by replacing weak test assertions with precise positive checks. No production code changes are required. **Finding C4-1** — `tests/server/test_tools_dispatch_halt.py`: Five tests that verify dispatch proceeds past the halt gate use `assert result.get("error") != "fleet_campaign_halted"`. Each test calls `_setup_standard_dispatch()` which wires a valid recipe and executor, so the expected outcome is that the dispatch proceeds past the halt gate — `assert "dispatch_id" in result` is the correct assertion. **Finding C4-3** — `tests/cli/test_doctor.py` lines 448 and 464: Two tests assert `checks[0]["severity"] in ("warning", "error")`. Source inspection of `_check_plugin_cache_exists` and `_check_installed_plugins_entry` confirms both return `Severity.WARNING` unconditionally under the test conditions. Closes #1887 ## Implementation Plan Plan file: `/home/talon/projects/autoskillit-runs/strengthen-assertions-20260505-211450-512860/.autoskillit/temp/make-plan/strengthen_assertions_plan_2026-05-05_211450.md` 🤖 Generated with [Claude Code](https://claude.com/claude-code) via AutoSkillit <!-- autoskillit:pipeline-signature steps=prepare_pr,run_arch_lenses,compose_pr,annotate_pr_diff,review_pr --> Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
## Summary Add a validate-only pre-commit hook (`scripts/check_sub_claude_md.py`) that checks every sub-CLAUDE.md file table mentions all `.py` files in its directory. This catches the gap at commit time (during `pre-commit run --all-files`) instead of at CI time, preventing the systematic 5+ CI round-trip failures observed since PR #1820. The script replicates the coverage logic from `test_sub_claude_md_covers_all_py_files` and `test_tests_sub_claude_md_covers_all_py_files`, using the same `EXPECTED_SUB_CLAUDE_MDS` lists. A new `.pre-commit-config.yaml` stanza triggers it on `.py` file changes under `src/autoskillit/` and `tests/`. ## Requirements - New script: `scripts/check_sub_claude_md.py` (validate-only, exits 1 with structured message on mismatch) - New stanza in `.pre-commit-config.yaml` triggered on `files: ^(tests/|src/autoskillit/).*\.py$` - Must check both `tests/<subdir>/CLAUDE.md` and `src/autoskillit/<subdir>/CLAUDE.md` file tables - Must use the same `EXPECTED_SUB_CLAUDE_MDS` lists as the test files (or derive from disk) - No auto-fix — the agent must manually add the row with a meaningful Purpose description - Pattern: `pass_filenames: false` (like existing `doc-counts` hook) ## Changed Files ### New (★): ★ scripts/check_sub_claude_md.py ★ tests/docs/test_check_sub_claude_md_script.py ### Modified (●): ● .pre-commit-config.yaml ● tests/docs/CLAUDE.md ● tests/infra/test_ci_dev_config.py Closes #1975 ## Implementation Plan Plan file: `/home/talon/projects/autoskillit-runs/impl-20260505-211945-173795/.autoskillit/temp/make-plan/add_pre_commit_hook_sub_claude_md_plan_2026-05-05_213000.md` 🤖 Generated with [Claude Code](https://claude.com/claude-code) via AutoSkillit <!-- autoskillit:pipeline-signature steps=prepare_pr,run_arch_lenses,compose_pr,annotate_pr_diff,review_pr --> --------- Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
## Summary Replace `test_config_resolution_fleet_enabled_via_experimental` in `tests/config/test_fleet_config.py` (lines 140–155) with an isolated version that uses `tmp_path` to create a synthetic `config.yaml` containing `experimental_enabled: true`, then calls `load_config(tmp_path)` to exercise the full config loading pipeline. The live-disk read (`Path(__file__).parents[2] / ".autoskillit" / "config.yaml"`) and its CI skip guard are removed entirely. The test is **rewritten** (not deleted) because `test_is_feature_enabled_fleet_defaults_false` (in `tests/core/test_type_constants.py`) only exercises `is_feature_enabled()` in isolation — it does not cover the `load_config()` → Dynaconf layer merge → `AutomationConfig.from_dynaconf()` → `is_feature_enabled()` pipeline. The rewrite provides genuine CI coverage of that end-to-end path, which the delete option would leave uncovered. ## Implementation Plan Plan file: `/home/talon/projects/autoskillit-runs/impl-20260505-221515-520014/.autoskillit/temp/make-plan/remove_live_config_read_plan_2026-05-05_221745.md` 🤖 Generated with [Claude Code](https://claude.com/claude-code) via AutoSkillit <!-- autoskillit:pipeline-signature steps=prepare_pr,run_arch_lenses,compose_pr,annotate_pr_diff,review_pr --> Co-authored-by: Trecek <trecek@users.noreply.github.com> Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
…1981) ## Summary C4-6 (load_config(tmp_path / "settings.toml") wrong arg type) was already resolved by a prior split commit. C4-9 (three separate load_config() calls for one-liner assertions in TestWorkspaceConfig) is the sole remaining change: consolidate three test methods into a single test_workspace_config_defaults method calling load_config(tmp_path) once and asserting all fields together. Drop the hasattr check as redundant. ## Requirements ### C4-6 — load_config() argument bug (ALREADY RESOLVED) Fix `load_config(tmp_path / "settings.toml")` → `load_config(tmp_path)`. Passing a `.toml` file path causes silent fallback to defaults. (Resolved in commit a22cb18.) ### C4-9 — Consolidate TestWorkspaceConfig assertions (REQUIRES ACTION) Consolidate `TestWorkspaceConfig`'s three separate `load_config(tmp_path)` calls into a single `test_workspace_config_defaults` test checking all three fields at once. Drop the `hasattr` check. ## Changed Files ### Modified (●): tests/config/test_config.py Closes #1888 ## Implementation Plan Plan file: `/home/talon/projects/autoskillit-runs/impl-20260505-221514-570647/.autoskillit/temp/make-plan/fix_load_config_argument_bug_and_consolidate_workspace_assertions_plan_2026-05-05_222100.md` 🤖 Generated with [Claude Code](https://claude.com/claude-code) via AutoSkillit <!-- autoskillit:pipeline-signature steps=prepare_pr,run_arch_lenses,compose_pr,annotate_pr_diff,review_pr --> --------- Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
## Summary Two test files contain stale docstrings and assertion messages that reference old names or misstate what the assertions actually check: - **C5-2** (`tests/workspace/test_skills.py`): `test_bundled_skills_list_matches_filesystem` docstring and failure message still say `make-script-skill` — the skill was renamed to `write-recipe`. - **C5-5** (`tests/execution/test_process_submodules.py`): per-symbol test docstrings say `"exports X"` but the assertions verify `__module__` (definition origin), not `__all__` membership. Each docstring should say `"is defined in X submodule"`. Both fixes are pure string edits in test files. No logic changes, no new fixtures, no isolation concerns. ## Implementation Plan Plan file: `/home/talon/projects/autoskillit-runs/impl-20260505-221516-303979/.autoskillit/temp/make-plan/fix_stale_docstrings_workspace_execution_tests_plan_2026-05-05_000000.md` 🤖 Generated with [Claude Code](https://claude.com/claude-code) via AutoSkillit <!-- autoskillit:pipeline-signature steps=prepare_pr,run_arch_lenses,compose_pr,annotate_pr_diff,review_pr --> --------- Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
…1979) ## Summary Create `src/autoskillit/recipes/research-archive.yaml` as a standalone sub-recipe that extracts the 9-step archival phase from `research.yaml` (lines 855–963). The critical change from the parent recipe: all ingredient-sourced values (`pr_url`, `worktree_path`, `research_dir`, `base_branch`) use `inputs.X` references instead of `context.X`, since these are declared ingredients in the standalone recipe rather than step-captured context variables. Step-captured values (`experiment_branch`, `artifact_branch`, `artifact_pr_url`, `archive_tag`) correctly remain as `context.X`. Pack declarations are `[github, ci]` (the archival phase only needs GitHub CLI and CI tools, not the full `research` pack). No `autoskillit_version` field — consistent with bundled recipe policy (removed in #1950). Only 4 ingredients: 3 campaign-sourced hidden (`worktree_path`, `research_dir`, `pr_url`) and 1 user-input (`base_branch`). The `report_path_after_finalize` and `source_dir` ingredients from the parent recipe are omitted because no archival step references them. Closes #1703 ## Implementation Plan Plan file: `/home/talon/projects/autoskillit-runs/impl-20260505-213323-163152/.autoskillit/temp/make-plan/p2_wp4_create_research_archive_yaml_sub_recipe_plan_2026-05-05_213600.md` 🤖 Generated with [Claude Code](https://claude.com/claude-code) via AutoSkillit <!-- autoskillit:pipeline-signature steps=prepare_pr,run_arch_lenses,compose_pr,annotate_pr_diff,review_pr --> --------- Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
## Summary The hook guard system lacks structural enforcement that command-inspecting guards must cover all tool variants that execute shell commands. Guards are written independently with ad-hoc extraction of command text, and the test suite validates each guard only against the tool format it was designed for — not against all tools it should logically intercept. The fix adds a structural meta-test that makes it impossible to register a command-inspecting guard without covering both the `Bash` native tool and `run_cmd` MCP tool, plus a parametrized test helper that forces every such guard to prove it blocks dangerous commands through either tool pathway. This closes a gap where `unsafe_install_guard.py` and `pr_create_guard.py` only read `tool_input.cmd` (from `run_cmd`) but ignore `tool_input.command` (from `Bash`), allowing headless agents to bypass these guards entirely when using the native Bash tool instead of the MCP wrapper. Closes #1980 ## Implementation Plan Plan file: `.autoskillit/temp/rectify/rectify_unsafe-install-guard-bash-tool-coverage-gap_2026-05-05_223500.md` 🤖 Generated with [Claude Code](https://claude.com/claude-code) via AutoSkillit <!-- autoskillit:pipeline-signature steps=prepare_pr,run_arch_lenses,compose_pr,annotate_pr_diff,review_pr --> --------- Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
Trecek
commented
May 8, 2026
Collaborator
Author
Trecek
left a comment
Trecek
left a comment
There was a problem hiding this comment.
AutoSkillit PR Review (individual)
| @@ -436,3 +450,129 @@ async def batch_cleanup_clones( | |||
| except Exception as exc: | |||
| logger.warning("batch_cleanup_clones failed", exc_info=True) | |||
| return json.dumps({"deleted": [], "preserved": [], "error": str(exc)}) | |||
Collaborator
Author
There was a problem hiding this comment.
[warning] defense: batch_cleanup_clones error handler returns result missing delete_failures key. Callers get KeyError.
Trecek
commented
May 8, 2026
Collaborator
Author
Trecek
left a comment
Trecek
left a comment
There was a problem hiding this comment.
AutoSkillit PR Review (individual)
| parts.append(b.get("text", "")) | ||
| # Non-text blocks (thinking, tool_use, etc.) contribute no text | ||
| self.result = "\n".join(parts) | ||
| elif not isinstance(self.result, str): |
Collaborator
Author
There was a problem hiding this comment.
[warning] slop: Unreachable elif branch in post_init: elif not isinstance(self.result, str) is dead code.
Trecek
commented
May 8, 2026
Collaborator
Author
Trecek
left a comment
Trecek
left a comment
There was a problem hiding this comment.
AutoSkillit PR Review (batch 4)
| description="L3 Fleet Orchestrator — multi-session campaign dispatch", | ||
| tool_tags=frozenset({"fleet"}), | ||
| skill_categories=frozenset({"fleet"}), | ||
| import_package="autoskillit.fleet", |
Collaborator
Author
There was a problem hiding this comment.
[critical] arch: IL-0 module stores import_package='autoskillit.fleet' and 'autoskillit.planner' as string data in FEATURE_REGISTRY. No actual import at module load. But importlib.import_module usage would violate layer boundary.
This finding requires a human decision — the correct path is ambiguous.
| ended_at = time.time() | ||
|
|
||
| # --- Timeout pre-check: short-circuit before result-block parsing --- | ||
| if skill_result.subtype == "timeout": |
Collaborator
Author
There was a problem hiding this comment.
[warning] bugs: skill_result.subtype compared as string literal == 'timeout'. Should use CliSubtype.TIMEOUT enum for type safety.
|
|
||
| try: | ||
| new_version = importlib.metadata.version("autoskillit") | ||
| except Exception: |
Collaborator
Author
There was a problem hiding this comment.
[warning] defense: _verify_update_result catches bare except Exception and falls back to new_version=current. Cannot distinguish version unchanged from infrastructure error.
|
|
||
| if Version(latest) > Version(current): | ||
| return Signal("binary", f"New release: {latest} (you have {current})") | ||
| except Exception: |
Collaborator
Author
There was a problem hiding this comment.
[warning] defense: _binary_signal catches bare except Exception and returns None. Missing packaging import silently returns no-signal.
| if not isinstance(conditions, list): | ||
| return False | ||
| return condition in conditions | ||
| except Exception: |
Collaborator
Author
There was a problem hiding this comment.
[warning] defense: _is_dismissed catches bare except Exception and returns False. Malformed dismissed_version causes repeated prompts.
| args=["autoskillit", "install"], returncode=0 | ||
| ) | ||
| with terminal_guard(): | ||
| subprocess.run(cmd, check=False, env=skip_env) |
Collaborator
Author
There was a problem hiding this comment.
[warning] defense: _run_update_sequence: upgrade subprocess return code not inspected. Failed upgrade silently ignored.
| extras: dict[str, str] = { | ||
| "AUTOSKILLIT_HEADLESS": "1", | ||
| "AUTOSKILLIT_SESSION_TYPE": SESSION_TYPE_SKILL, | ||
| "MAX_MCP_OUTPUT_TOKENS": _MAX_MCP_OUTPUT_TOKENS_VALUE, |
Collaborator
Author
There was a problem hiding this comment.
[warning] defense: MAX_MCP_OUTPUT_TOKENS and MCP_CONNECTION_NONBLOCKING duplicated between extras dict and _SESSION_BASELINE_ENV. Silent maintenance hazard if canonical values change.
| extras[KITCHEN_SESSION_ID_ENV_VAR] = kitchen_session_id | ||
| if allowed_write_prefix: | ||
| extras["AUTOSKILLIT_ALLOWED_WRITE_PREFIX"] = allowed_write_prefix | ||
| # Layer caller env_extras (campaign vars) UNDER the mandatory keys. |
Collaborator
Author
There was a problem hiding this comment.
[warning] defense: build_food_truck_cmd: caller env_extras keys silently dropped if they match mandatory keys. No warning logged.
| f"Recipe '{recipe}' could not be loaded: {exc}", | ||
| ) | ||
|
|
||
| _DISPATCHABLE_KINDS = frozenset({"standard", "food-truck"}) |
Collaborator
Author
There was a problem hiding this comment.
[warning] defense: _DISPATCHABLE_KINDS defined inside _run_dispatch (recreated every call). Should be module-level constant.
| except Exception as exc: | ||
| logger.error("run_skill unhandled exception", exc_info=True) | ||
| return SkillResult.crashed( | ||
| exception=exc, | ||
| skill_command=skill_command, | ||
| order_id=order_id, | ||
| ).to_json() | ||
| except BaseException: |
Collaborator
Author
There was a problem hiding this comment.
[warning] defense: CancelledError (BaseException) raised inside inner try falls through to except Exception and returns crashed SkillResult rather than being re-raised.
This finding requires a human decision — the correct path is ambiguous.
Trecek
commented
May 8, 2026
Collaborator
Author
Trecek
left a comment
Trecek
left a comment
There was a problem hiding this comment.
AutoSkillit PR Review (individual)
| if new_version != current: | ||
| return True | ||
|
|
||
| from autoskillit.cli._install_info import upgrade_command |
Collaborator
Author
There was a problem hiding this comment.
[warning] slop: Redundant import of upgrade_command inside _verify_update_result — already imported at module level.
Trecek
commented
May 8, 2026
Collaborator
Author
Trecek
left a comment
Trecek
left a comment
There was a problem hiding this comment.
AutoSkillit PR Review (individual)
| except Exception as exc: | ||
| logger.warning("load_recipe failed for '%s'", recipe, exc_info=True) | ||
| return fleet_error( | ||
| FleetErrorCode.FLEET_RECIPE_NOT_FOUND, |
Collaborator
Author
There was a problem hiding this comment.
[warning] slop: _DISPATCHABLE_KINDS defined as frozenset inside _run_dispatch on every call. Should be module-level constant.
Trecek
commented
May 8, 2026
Collaborator
Author
Trecek
left a comment
Trecek
left a comment
There was a problem hiding this comment.
AutoSkillit PR Review (individual)
| @@ -170,7 +183,19 @@ async def run_python( | |||
| return json.dumps({"success": False, "error": f"{type(exc).__name__}: {exc}"}) | |||
|
|
|||
|
|
|||
| @mcp.tool(tags={"autoskillit", "kitchen"}, annotations={"readOnlyHint": True}) | |||
| def _persist_run_skill_state(skill_result: SkillResult, project_dir: Path) -> None: | |||
Collaborator
Author
There was a problem hiding this comment.
[warning] slop: _persist_run_skill_state and _clear_run_skill_state are unnecessary single-line wrapper functions. Deferred imports could be inlined.
Trecek
commented
May 8, 2026
Collaborator
Author
There was a problem hiding this comment.
(L403 — outside diff hunk) [warning] bugs: _check_merge_base_unpublished accesses step.on_result.routes.values() without guard. If routes is None, raises AttributeError.
Trecek
commented
May 8, 2026
Collaborator
Author
There was a problem hiding this comment.
(L367 — outside diff hunk) [warning] defense: output_mode validation for research recipe is hardcoded name-based special case in generic open_kitchen handler. Should be driven by recipe schema.
This finding requires a human decision — the correct path is ambiguous.
Trecek
commented
May 8, 2026
Collaborator
Author
There was a problem hiding this comment.
(L275 — outside diff hunk) [warning] slop: _check_always_has_no_write_exit has multi-line docstring whose first sentence repeats the @semantic_rule description exactly.
Trecek
commented
May 8, 2026
Collaborator
Author
There was a problem hiding this comment.
(L270 — outside diff hunk) [warning] slop: install_result initialized to dummy CompletedProcess only to satisfy type checker before with block overwrites it. Use Optional typing instead.
Trecek
commented
May 8, 2026
Collaborator
Author
Trecek
left a comment
Trecek
left a comment
There was a problem hiding this comment.
AutoSkillit PR Review — Verdict: approved_with_comments
Scope: Top 20 most-changed Python source files out of 1,551 total changed files (287K LoC added).
37 warning-level findings across 15 files. No blocking changes required. See inline comments.
Finding Summary by Dimension
| Dimension | Findings | Key Pattern |
|---|---|---|
| defense | 17 | Broad except Exception handlers swallowing errors silently (headless flush, update checks, rule loaders) |
| bugs | 9 | Assert-as-validation (disabled under -O), string-vs-enum comparisons, token log ordering |
| slop | 8 | Dead code branches, redundant imports, unnecessary wrapper functions, duplicate logic |
| cohesion | 2 | Asymmetric SkillResult construction across recovery paths |
| arch | 1 | IL-0 FEATURE_REGISTRY storing IL-2 import paths as string data |
Top Patterns
-
Silent exception swallowing (10 findings):
except Exception: passorexcept Exception: return []patterns suppress infrastructure failures. Most common inheadless/__init__.py(flush paths),_update_checks.py, and recipe rule loaders. Recommended: narrow exception types or raise log level to WARNING. -
Assert-as-validation (2 findings):
assert x is not Noneused for runtime invariants that should beif x is None: raise RuntimeError(...)— asserts are stripped under-O. -
Dead code / unnecessary wrappers (5 findings): Unreachable branches (
_session_model.py:82), redundant one-line delegation functions (_cmd_rpc.py:127,170), duplicate logic (tools_git.py:24). -
Type safety (2 findings): String literal comparison
== "timeout"instead of enumCliSubtype.TIMEOUTinfleet/_api.py.
Files NOT Reviewed
This review covers only the top 20 source files by addition count. Notable files NOT reviewed include:
- All test files (1,400+ lines of test changes)
- Recipe YAML/JSON contracts
- Skills and SKILL.md files
- Configuration, documentation, and CI files
- The remaining ~1,500 changed files
## Summary Add `slots=True` to all 63 `@dataclass(frozen=True)` definitions (across 33 files in 11 packages) that currently lack it. Python `slots=True` on a frozen dataclass eliminates the `__dict__` per-instance overhead and replaces it with typed slot descriptors, reducing memory usage and improving attribute access speed. The project's `requires-python = ">=3.11"` means `slots=True` on `@dataclass` (introduced in 3.10) is fully supported. Two frozen dataclasses already have `slots=True` (`result_parser.py`, `_hook_settings.py`) and serve as the established pattern for this change. Every affected dataclass was verified to be safe: no inheritance chains between dataclasses, no manually defined `__slots__`, and no non-dataclass parent that would conflict. A new architecture compliance test is introduced first to establish the invariant and prevent future regressions. Closes #2192 ## Implementation Plan Plan file: `/home/talon/projects/autoskillit-runs/impl-20260507-215724-550140/.autoskillit/temp/make-plan/perf_add_slots_true_frozen_dataclasses_plan_2026-05-07_000001.md` 🤖 Generated with [Claude Code](https://claude.com/claude-code) via AutoSkillit <!-- autoskillit:pipeline-signature steps=prepare_pr,run_arch_lenses,compose_pr,annotate_pr_diff,review_pr --> ## Token Usage Summary | Step | Model | count | uncached | output | cache_read | peak_ctx | turns | cache_write | time | |------|-------|-------|----------|--------|------------|----------|-------|-------------|------| | plan | claude-sonnet-4-6 | 1 | 115 | 10.6k | 396.2k | 45.2k | 82 | 51.2k | 6m 58s | | verify | claude-sonnet-4-6 | 1 | 204 | 21.8k | 1.2M | 59.7k | 66 | 48.2k | 5m 3s | | implement* | MiniMax-M2.7-highspeed | 1 | 1.1M | 15.0k | 1.0M | 35.1k | 166 | 52.5k | 4m 45s | | fix | claude-sonnet-4-6 | 1 | 360 | 17.2k | 2.7M | 90.0k | 108 | 79.7k | 15m 58s | | prepare_pr* | MiniMax-M2.7-highspeed | 1 | 58.8k | 4.2k | 149.0k | 29.8k | 16 | 42.3k | 1m 26s | | compose_pr* | MiniMax-M2.7-highspeed | 1 | 49.7k | 1.5k | 175.9k | 29.8k | 15 | 15.0k | 50s | | **Total** | | | 1.3M | 70.3k | 5.7M | 90.0k | | 289.0k | 35m 1s | \* *Step used a non-Anthropic provider; caching behavior may differ.* ## Token Efficiency | Step | LoC Changed | cache_read/LoC | cache_write/LoC | output/LoC | |------|-------------|----------------|-----------------|------------| | plan | 0 | — | — | — | | verify | 0 | — | — | — | | implement | 183 | 5727.6 | 287.1 | 81.7 | | fix | 16 | 170299.2 | 4982.6 | 1077.5 | | prepare_pr | 0 | — | — | — | | compose_pr | 0 | — | — | — | | **Total** | **199** | 28542.7 | 1452.1 | 353.0 | ## Model Usage Breakdown | Model | steps | uncached | output | cache_read | cache_write | time | |-------|-------|----------|--------|------------|-------------|------| | claude-sonnet-4-6 | 3 | 679 | 49.6k | 4.3M | 179.1k | 27m 59s | | MiniMax-M2.7-highspeed | 3 | 1.3M | 20.6k | 1.4M | 109.9k | 7m 1s | --------- Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
…#2217) ## Summary Replace `return asdict(self)` with explicit field dicts in the `to_dict()` method of four hot-path dataclasses: `DispatchRecord`, `TokenEntry`, `TimingEntry`, and `FailureRecord`. For `DispatchRecord.token_usage` (the single non-primitive field across all four), use `dict(self.token_usage)` for a shallow copy that avoids the deep-copy overhead of `asdict` while preserving safety for current callers. Remove `asdict` from the `dataclasses` import in all four files once it is no longer used. The cold-path `AutomationConfig` serialization at `cli/app.py:302` is explicitly out of scope and must remain unchanged. ## Requirements ## Acceptance Criteria - [ ] All 4 warm-path `to_dict()` methods use explicit field dicts - [ ] `DispatchRecord.token_usage` uses shallow copy (`dict(self.token_usage)`) - [ ] `AutomationConfig` at `cli/app.py:302` keeps `asdict()` - [ ] All existing tests pass - [ ] JSON output is identical (verified by round-trip tests) ## Changed Files ### New (★): tests/fleet/test_state_schema.py tests/pipeline/test_audit.py ### Modified (●): src/autoskillit/core/types/_type_results.py src/autoskillit/fleet/state_types.py src/autoskillit/pipeline/timings.py src/autoskillit/pipeline/tokens.py Closes #2193 ## Implementation Plan Plan file: `/home/talon/projects/autoskillit-runs/impl-20260507-215724-984456/.autoskillit/temp/make-plan/perf_replace_asdict_plan_2026-05-07_220015.md` 🤖 Generated with [Claude Code](https://claude.com/claude-code) via AutoSkillit <!-- autoskillit:pipeline-signature steps=prepare_pr,run_arch_lenses,compose_pr,annotate_pr_diff,review_pr --> ## Token Usage Summary | Step | Model | count | uncached | output | cache_read | peak_ctx | turns | cache_write | time | |------|-------|-------|----------|--------|------------|----------|-------|-------------|------| | plan | claude-sonnet-4-6 | 1 | 107 | 10.0k | 340.2k | 61.7k | 40 | 32.2k | 4m 19s | | verify | claude-sonnet-4-6 | 1 | 68 | 8.6k | 256.7k | 46.2k | 42 | 33.3k | 4m 32s | | implement* | MiniMax-M2.7-highspeed | 1 | 1.2M | 6.5k | 831.6k | 29.8k | 77 | 16.3k | 2m 48s | | prepare_pr* | MiniMax-M2.7-highspeed | 1 | 73.3k | 3.7k | 205.7k | 29.8k | 20 | 15.3k | 1m 34s | | compose_pr* | MiniMax-M2.7-highspeed | 1 | 28.6k | 1.5k | 146.1k | 29.8k | 14 | 15.1k | 45s | | **Total** | | | 1.3M | 30.5k | 1.8M | 61.7k | | 112.2k | 14m 1s | \* *Step used a non-Anthropic provider; caching behavior may differ.* ## Token Efficiency | Step | LoC Changed | cache_read/LoC | cache_write/LoC | output/LoC | |------|-------------|----------------|-----------------|------------| | plan | 0 | — | — | — | | verify | 0 | — | — | — | | implement | 136 | 6114.4 | 119.6 | 48.1 | | prepare_pr | 0 | — | — | — | | compose_pr | 0 | — | — | — | | **Total** | **136** | 13090.7 | 825.1 | 223.9 | ## Model Usage Breakdown | Model | steps | uncached | output | cache_read | cache_write | time | |-------|-------|----------|--------|------------|-------------|------| | claude-sonnet-4-6 | 2 | 175 | 18.7k | 596.9k | 65.5k | 8m 52s | | MiniMax-M2.7-highspeed | 3 | 1.3M | 11.8k | 1.2M | 46.7k | 5m 8s | --------- Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
## Summary
Change validate-audit (project-local), validate-test-audit, and the
extended validate-audit SKILL.md files from writing flat timestamped
files into a shared `validate-audit/` directory to creating a per-run
timestamped subdirectory (`validate-audit-{YYYY-MM-DD_HHMMSS}/`) with
timestamp-free filenames inside. This matches the established pattern
used by `validate-team`. Update downstream path references in
`skill_contracts.yaml` and all affected tests.
Closes #1960
## Implementation Plan
Plan file:
`/home/talon/projects/autoskillit-runs/impl-20260507-225440-809232/.autoskillit/temp/make-plan/validate-audit_adopt_per-run_subdirectory_output_pattern_plan_2026-05-07_225500.md`
🤖 Generated with [Claude Code](https://claude.com/claude-code) via
AutoSkillit
<!-- autoskillit:pipeline-signature
steps=prepare_pr,run_arch_lenses,compose_pr,annotate_pr_diff,review_pr
-->
## Token Usage Summary
| Step | Model | count | uncached | output | cache_read | peak_ctx |
turns | cache_write | time |
|------|-------|-------|----------|--------|------------|----------|-------|-------------|------|
| build_execution_map | claude-sonnet-4-6 | 1 | 76 | 12.2k | 347.2k |
52.1k | 31 | 39.9k | 4m 22s |
| plan | claude-opus-4-6 | 1 | 79 | 19.2k | 755.8k | 79.2k | 57 | 85.7k
| 7m 20s |
| verify | claude-sonnet-4-6 | 1 | 29 | 15.7k | 447.8k | 54.6k | 77 |
41.5k | 7m 44s |
| implement* | MiniMax-M2.7-highspeed | 1 | 3.8M | 20.0k | 2.3M | 29.8k
| 178 | 72.2k | 7m 30s |
| fix | claude-sonnet-4-6 | 1 | 190 | 10.7k | 1.1M | 60.6k | 61 | 50.3k
| 8m 29s |
| prepare_pr* | MiniMax-M2.7-highspeed | 1 | 89.7k | 4.2k | 208.5k |
29.8k | 20 | 42.2k | 1m 32s |
| compose_pr* | MiniMax-M2.7-highspeed | 1 | 68.6k | 1.8k | 172.3k |
28.7k | 16 | 41.0k | 55s |
| **Total** | | | 3.9M | 83.7k | 5.3M | 79.2k | | 372.7k | 37m 54s |
\* *Step used a non-Anthropic provider; caching behavior may differ.*
## Token Efficiency
| Step | LoC Changed | cache_read/LoC | cache_write/LoC | output/LoC |
|------|-------------|----------------|-----------------|------------|
| build_execution_map | 0 | — | — | — |
| plan | 0 | — | — | — |
| verify | 0 | — | — | — |
| implement | 162 | 13957.0 | 446.0 | 123.2 |
| fix | 2 | 532814.5 | 25159.0 | 5347.5 |
| prepare_pr | 0 | — | — | — |
| compose_pr | 0 | — | — | — |
| **Total** | **164** | 32062.6 | 2272.9 | 510.5 |
## Model Usage Breakdown
| Model | steps | uncached | output | cache_read | cache_write | time |
|-------|-------|----------|--------|------------|-------------|------|
| claude-sonnet-4-6 | 3 | 295 | 38.6k | 1.9M | 131.6k | 20m 35s |
| claude-opus-4-6 | 1 | 79 | 19.2k | 755.8k | 85.7k | 7m 20s |
| MiniMax-M2.7-highspeed | 3 | 3.9M | 25.9k | 2.6M | 155.4k | 9m 57s |
---------
Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
## Summary The recipe validation pipeline has an asymmetric structural gap: the feature-gate axis has a complete `undeclared-feature-requirement` rule (ERROR severity) that statically cross-references every `run_skill` step's categories against `FEATURE_REGISTRY`, but the pack-gate axis has no equivalent rule. This gap caused three incidents over five weeks. The fix adds an `undeclared-pack-requirement` semantic rule mirroring the feature-gate pattern, makes `unknown-required-pack` ERROR-severity, fixes `research-design.yaml` to declare `vis-lens`, and updates the test that locked in the incorrect `requires_packs` value. ## Requirements ## Conflict Resolution Decisions The following files had merge conflicts that were automatically resolved. ## Changed Files ### Modified (●): - `src/autoskillit/recipe/rules/rules_packs.py` - `src/autoskillit/recipe/rules/rules_skills.py` - `src/autoskillit/recipes/research-design.json` - `src/autoskillit/recipes/research-design.yaml` - `tests/recipe/test_bundled_recipes_general.py` - `tests/recipe/test_bundled_recipes_research_design.py` - `tests/recipe/test_rules_packs.py` Closes #2220 ## Implementation Plan Plan file: `/home/talon/projects/autoskillit-runs/remediation-20260507-235311-480573/.autoskillit/temp/rectify/rectify_undeclared_pack_requirement_2026-05-08_000100.md` 🤖 Generated with [Claude Code](https://claude.com/claude-code) via AutoSkillit <!-- autoskillit:pipeline-signature steps=prepare_pr,run_arch_lenses,compose_pr,annotate_pr_diff,review_pr --> ## Token Usage Summary | Step | Model | count | uncached | output | cache_read | peak_ctx | turns | cache_write | time | |------|-------|-------|----------|--------|------------|----------|-------|-------------|------| | rectify | claude-sonnet-4-6 | 1 | 6.6k | 13.1k | 907.1k | 118.7k | 197 | 67.7k | 8m 41s | | dry_walkthrough | claude-sonnet-4-6 | 1 | 40 | 9.6k | 547.7k | 53.1k | 99 | 40.0k | 4m 56s | | implement* | MiniMax-M2.7-highspeed | 1 | 1.9M | 16.7k | 1.6M | 70.4k | 160 | 129.5k | 7m 54s | | assess | claude-opus-4-6 | 1 | 84 | 9.4k | 2.1M | 71.5k | 88 | 60.9k | 6m 41s | | prepare_pr* | MiniMax-M2.7-highspeed | 1 | 83.5k | 2.9k | 206.4k | 34.0k | 22 | 52.2k | 1m 25s | | compose_pr* | MiniMax-M2.7-highspeed | 1 | 44.4k | 1.2k | 198.2k | 28.7k | 14 | 15.0k | 40s | | **Total** | | | 2.1M | 52.9k | 5.5M | 118.7k | | 365.3k | 30m 20s | \* *Step used a non-Anthropic provider; caching behavior may differ.* ## Token Efficiency | Step | LoC Changed | cache_read/LoC | cache_write/LoC | output/LoC | |------|-------------|----------------|-----------------|------------| | rectify | 0 | — | — | — | | dry_walkthrough | 0 | — | — | — | | implement | 234 | 6951.9 | 553.6 | 71.2 | | assess | 21 | 98049.5 | 2898.2 | 446.5 | | prepare_pr | 0 | — | — | — | | compose_pr | 0 | — | — | — | | **Total** | **255** | 21745.4 | 1432.6 | 207.3 | ## Model Usage Breakdown | Model | steps | uncached | output | cache_read | cache_write | time | |-------|-------|----------|--------|------------|-------------|------| | claude-sonnet-4-6 | 2 | 6.7k | 22.7k | 1.5M | 107.7k | 13m 38s | | MiniMax-M2.7-highspeed | 3 | 2.1M | 20.8k | 2.0M | 196.7k | 9m 59s | | claude-opus-4-6 | 1 | 84 | 9.4k | 2.1M | 60.9k | 6m 41s | --------- Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
## Summary The `fleet campaign` session launches without an `initial_message`, so Claude has no first user turn to trigger the `FIRST ACTION` block that displays the ingredient table. The `fleet dispatch` path and `order` path both correctly pass a greeting as `initial_message`. The fix adds a `_FLEET_CAMPAIGN_GREETINGS` list and wires it through the campaign launch path, mirroring the existing dispatch pattern. Closes #2214 ## Implementation Plan Plan file: `/home/talon/projects/autoskillit-runs/impl-20260508-010012-804577/.autoskillit/temp/make-plan/fleet_campaign_session_missing_initial_message_greeting_trigger_plan_2026-05-08_010012.md` 🤖 Generated with [Claude Code](https://claude.com/claude-code) via AutoSkillit <!-- autoskillit:pipeline-signature steps=prepare_pr,run_arch_lenses,compose_pr,annotate_pr_diff,review_pr --> ## Token Usage Summary | Step | Model | count | uncached | output | cache_read | peak_ctx | turns | cache_write | time | |------|-------|-------|----------|--------|------------|----------|-------|-------------|------| | plan | claude-opus-4-6 | 1 | 75 | 9.1k | 933.1k | 62.0k | 71 | 52.3k | 5m 8s | | verify | claude-opus-4-6 | 1 | 37 | 7.3k | 734.2k | 46.7k | 59 | 33.5k | 4m 3s | | implement* | MiniMax-M2.7-highspeed | 1 | 1.1M | 12.7k | 1.0M | 34.0k | 100 | 56.4k | 5m 43s | | fix | claude-sonnet-4-6 | 1 | 374 | 30.8k | 3.4M | 105.9k | 118 | 92.9k | 11m 35s | | prepare_pr* | MiniMax-M2.7-highspeed | 1 | 169.7k | 4.1k | 402.1k | 28.7k | 37 | 41.2k | 1m 56s | | compose_pr* | MiniMax-M2.7-highspeed | 1 | 52.2k | 1.2k | 227.0k | 28.7k | 16 | 15.1k | 42s | | **Total** | | | 1.3M | 65.4k | 6.7M | 105.9k | | 291.5k | 29m 9s | \* *Step used a non-Anthropic provider; caching behavior may differ.* ## Token Efficiency | Step | LoC Changed | cache_read/LoC | cache_write/LoC | output/LoC | |------|-------------|----------------|-----------------|------------| | plan | 0 | — | — | — | | verify | 0 | — | — | — | | implement | 153 | 6606.7 | 368.5 | 83.3 | | fix | 20 | 168857.8 | 4646.4 | 1541.3 | | prepare_pr | 0 | — | — | — | | compose_pr | 0 | — | — | — | | **Total** | **173** | 38638.6 | 1684.9 | 378.0 | ## Model Usage Breakdown | Model | steps | uncached | output | cache_read | cache_write | time | |-------|-------|----------|--------|------------|-------------|------| | claude-opus-4-6 | 2 | 112 | 16.5k | 1.7M | 85.8k | 9m 11s | | MiniMax-M2.7-highspeed | 3 | 1.3M | 18.1k | 1.6M | 112.8k | 8m 22s | | claude-sonnet-4-6 | 1 | 374 | 30.8k | 3.4M | 92.9k | 11m 35s | --------- Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
…2222) ## Summary The `research-campaign.yaml` declares 8 campaign-level ingredients but only forwards a subset to each dispatch's `ingredients:` block. Several ingredients (`task`, `review_design`, `output_mode`, `review_pr`, `audit_claims`) that sub-recipes declare are never forwarded — they silently fall back to sub-recipe defaults (or to the `_run_dispatch` auto-injection in the case of `task`), ignoring the user's campaign-level values. This plan fixes the YAML forwarding gaps and adds a new static validation rule (`campaign-dangling-ingredient`) that catches this class of bug at authoring time. ## Requirements (Embedded in the issue body — issue #2215 describes the problem and solution in detail) ## Implementation Plan Plan file: `/home/talon/projects/autoskillit-runs/impl-20260508-010013-712390/.autoskillit/temp/make-plan/research_campaign_ingredient_forwarding_plan_2026-05-08_010600.md` ## Changed Files ### Modified (●): ● src/autoskillit/recipe/rules/rules_campaign.py ● src/autoskillit/recipes/campaigns/research-campaign.json ● src/autoskillit/recipes/campaigns/research-campaign.yaml ● tests/recipe/test_campaign_loader.py ● tests/recipe/test_research_campaign_rules.py ● tests/recipe/test_rules_campaign.py Closes #2215 🤖 Generated with [Claude Code](https://claude.com/claude-code) via AutoSkillit <!-- autoskillit:pipeline-signature steps=prepare_pr,run_arch_lenses,compose_pr,annotate_pr_diff,review_pr --> ## Token Usage Summary | Step | Model | count | uncached | output | cache_read | peak_ctx | turns | cache_write | time | |------|-------|-------|----------|--------|------------|----------|-------|-------------|------| | plan | claude-opus-4-6 | 1 | 78 | 13.1k | 649.7k | 68.6k | 78 | 59.9k | 6m 11s | | verify | claude-opus-4-6 | 1 | 897 | 6.2k | 625.9k | 70.5k | 81 | 57.3k | 4m 27s | | implement* | MiniMax-M2.7-highspeed | 1 | 1.1M | 10.7k | 974.1k | 28.7k | 83 | 51.5k | 7m 50s | | fix | claude-sonnet-4-6 | 1 | 174 | 8.1k | 929.2k | 61.4k | 53 | 48.3k | 3m 47s | | prepare_pr* | MiniMax-M2.7-highspeed | 1 | 133.1k | 3.0k | 341.7k | 28.7k | 24 | 15.2k | 1m 21s | | compose_pr* | MiniMax-M2.7-highspeed | 1 | 37.0k | 1.4k | 169.5k | 28.7k | 14 | 15.1k | 44s | | **Total** | | | 1.2M | 42.4k | 3.7M | 70.5k | | 247.4k | 24m 21s | \* *Step used a non-Anthropic provider; caching behavior may differ.* ## Token Efficiency | Step | LoC Changed | cache_read/LoC | cache_write/LoC | output/LoC | |------|-------------|----------------|-----------------|------------| | plan | 0 | — | — | — | | verify | 0 | — | — | — | | implement | 297 | 3279.6 | 173.5 | 35.9 | | fix | 33 | 28156.7 | 1463.2 | 246.1 | | prepare_pr | 0 | — | — | — | | compose_pr | 0 | — | — | — | | **Total** | **330** | 11181.9 | 749.7 | 128.6 | ## Model Usage Breakdown | Model | steps | uncached | output | cache_read | cache_write | time | |-------|-------|----------|--------|------------|-------------|------| | claude-opus-4-6 | 2 | 975 | 19.3k | 1.3M | 117.2k | 10m 38s | | MiniMax-M2.7-highspeed | 3 | 1.2M | 15.0k | 1.5M | 81.9k | 9m 55s | | claude-sonnet-4-6 | 1 | 174 | 8.1k | 929.2k | 48.3k | 3m 47s | --------- Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
## Summary
Create a new `validate-review-decisions` skill at
`skills_extended/validate-review-decisions/SKILL.md` that adds mandatory
intent analysis and seven evidence-gathering rules to the validation
workflow for review-decisions audit reports. Update `full-audit.yaml` to
route review-decisions validation through this new skill instead of the
generic `validate-audit`. Add contract tests and update all test
manifests.
The skill follows the architecture established by `validate-test-audit`
(domain-specific semantic rules + intent analysis) while preserving full
output compatibility with `validate-audit` (same directory, naming
convention, `validated: true` sentinel, `AUTOSKILLIT_AUDIT_RUN_DIR`
support).
## Requirements
### REQ-VRD-1: Skill structure
The skill MUST be placed at
`skills_extended/validate-review-decisions/SKILL.md` with `categories:
[audit]` frontmatter.
### REQ-VRD-2: Intent analysis as mandatory step
Code validation subagents MUST perform intent analysis (docstring check,
git provenance, test coverage, contract analysis, architectural
constraint check, behavioral simulation) before assigning a verdict to
ANY finding. This is not optional — every finding must have an intent
analysis section in the subagent's reasoning.
### REQ-VRD-3: Evidence-gathering rules in subagent instructions
The skill MUST include the seven evidence-gathering rules in the code
validation subagent prompt. Rules MUST be generalizable (no references
to specific finding IDs).
### REQ-VRD-4: Output compatibility
Output files MUST use the same directory, naming convention, and format
as `validate-audit`.
### REQ-VRD-5: Standalone invocability
The skill MUST accept an `{audit_report_path}` argument and
auto-discover the most recent review-decisions audit report when
omitted.
### REQ-VRD-6: Full-audit recipe routing
Update `full-audit.yaml` to dispatch review-decisions audit validation
to `validate-review-decisions` and all other audit types to
`validate-audit` (or `validate-test-audit` for tests).
### REQ-VRD-7: No pack changes
The skill MUST use the existing `audit` pack.
### REQ-VRD-8: Consider generalizing intent analysis to validate-audit
(follow-up)
After `validate-review-decisions` and `validate-test-audit` both
establish intent analysis patterns, evaluate merging the common intent
analysis rules back into the generic `validate-audit` skill. This is a
follow-up, not a blocker.
## Implementation Plan
Plan file:
`/home/talon/projects/autoskillit-runs/impl-20260508-010014-405487/.autoskillit/temp/make-plan/validate_review_decisions_skill_plan_2026-05-08_010600.md`
🤖 Generated with [Claude Code](https://claude.com/claude-code) via
AutoSkillit
<!-- autoskillit:pipeline-signature
steps=prepare_pr,run_arch_lenses,compose_pr,annotate_pr_diff,review_pr
-->
## Token Usage Summary
| Step | Model | count | uncached | output | cache_read | peak_ctx |
turns | cache_write | time |
|------|-------|-------|----------|--------|------------|----------|-------|-------------|------|
| plan | claude-opus-4-6 | 1 | 2.6k | 14.9k | 1.8M | 93.1k | 105 | 79.9k
| 9m 1s |
| verify | claude-opus-4-6 | 1 | 1.7k | 11.6k | 1.4M | 63.3k | 116 |
50.4k | 6m 22s |
| implement* | MiniMax-M2.7-highspeed | 1 | 3.6M | 22.5k | 1.9M | 28.7k
| 180 | 16.2k | 12m 59s |
| fix | claude-sonnet-4-6 | 1 | 174 | 7.5k | 874.3k | 53.2k | 47 | 42.9k
| 3m 33s |
| prepare_pr* | MiniMax-M2.7-highspeed | 1 | 102.6k | 4.2k | 261.0k |
34.0k | 24 | 27.4k | 1m 37s |
| compose_pr* | MiniMax-M2.7-highspeed | 1 | 57.5k | 1.8k | 227.0k |
28.7k | 16 | 15.1k | 49s |
| **Total** | | | 3.7M | 62.5k | 6.4M | 93.1k | | 232.0k | 34m 23s |
\* *Step used a non-Anthropic provider; caching behavior may differ.*
## Token Efficiency
| Step | LoC Changed | cache_read/LoC | cache_write/LoC | output/LoC |
|------|-------------|----------------|-----------------|------------|
| plan | 0 | — | — | — |
| verify | 0 | — | — | — |
| implement | 766 | 2508.8 | 21.2 | 29.4 |
| fix | 2 | 437171.0 | 21471.5 | 3770.5 |
| prepare_pr | 0 | — | — | — |
| compose_pr | 0 | — | — | — |
| **Total** | **768** | 8375.0 | 302.1 | 81.3 |
## Model Usage Breakdown
| Model | steps | uncached | output | cache_read | cache_write | time |
|-------|-------|----------|--------|------------|-------------|------|
| claude-opus-4-6 | 2 | 4.3k | 26.5k | 3.1M | 130.2k | 15m 24s |
| MiniMax-M2.7-highspeed | 3 | 3.7M | 28.5k | 2.4M | 58.8k | 15m 25s |
| claude-sonnet-4-6 | 1 | 174 | 7.5k | 874.3k | 42.9k | 3m 33s |
---------
Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
…queued` State (#2225) ## Summary Formalize the issue label lifecycle as a first-class discrete system by introducing an `IssueLabelState` enum, a `LabelDef` registry with per-state metadata (color, description, swap semantics), and a transition table with validation — modeled after the fleet `DispatchStatus` pattern. Add `queued` as the fifth lifecycle state. Refactor `claim_issue`, `release_issue`, and `claim_and_resolve_issue` to derive colors, descriptions, and swap-remove sets from the registry instead of hardcoding them. Update `process-issues` Phase 0.5 to apply `queued` (not `in-progress`) during upfront claiming, and add a `queued → in-progress` swap at recipe pickup. ## Requirements ### R1: Formalize Label Lifecycle as a Discrete System Create an architectural component (enum + registry or Protocol-based approach) where each **lifecycle label** (as distinct from classification labels like `recipe:implementation`) is a first-class entity that declares: 1. **Label name** — the GitHub label string (e.g., `"queued"`, `"in-progress"`, `"staged"`, `"fail"`) 2. **Color** — hex color for `ensure_label` (currently hardcoded: `fbca04` for in-progress, `0075ca` for staged, `d73a4a` for fail) 3. **Description** — human-readable description for `ensure_label` 4. **Swap semantics** — which labels to remove when entering this state (e.g., entering `in-progress` removes `queued` and `fail`) 5. **Valid transitions** — which states this label can transition to (modeled after `_ALLOWED_TRANSITIONS` in fleet) When a new lifecycle label is added, the system must enforce that all of these are defined. No hanging labels — every label in the lifecycle participates in the tracking history. ### R2: Add `queued` Lifecycle State Add a `queued` label with the following lifecycle position: ``` [unlabeled / triaged] │ │ upfront claim (Phase 0.5) ▼ [queued] ← claimed by orchestrator, not yet processing │ │ recipe session begins ▼ [in-progress] ← recipe actively executing │ ├─ success → [staged] ├─ failure → [fail] └─ bare → [unlabeled] ``` Transitions for `queued`: - **Entry**: from unlabeled/fail (upfront claim swaps `fail` → `queued`) - **Exit**: to `in-progress` (recipe pickup), or to unlabeled (fatal cleanup release) ### R3: Integrate with `process-issues` Workflow Update `process-issues/SKILL.md` Phase 0.5 and Step 3: - **Phase 0.5 (upfront claiming)**: `claim_issue` should apply `queued` instead of `in-progress` - **Step 3b.1 (recipe pickup)**: Before loading the recipe, swap `queued` → `in-progress` for the specific issue being processed - **Fatal failure cleanup**: Release all `queued` issues (not just `in-progress`) back to unlabeled - **Recipe `claim_and_resolve` step**: When `upfront_claimed=true`, the issue arrives with `queued` (not `in-progress`), so the reentry path needs to handle the `queued` → `in-progress` swap ### R4: Integration Points These components need updates: | Component | File | Change | |-----------|------|--------| | `GitHubConfig` | `config/_config_dataclasses.py` | Add `queued_label` field (or migrate to registry) | | `defaults.yaml` | `config/defaults.yaml` | Add `queued` to `allowed_labels` and label config | | `claim_issue` | `server/tools/tools_issue_lifecycle.py` | Support claiming with `queued` label | | `claim_and_resolve_issue` | `server/tools/tools_issue_composite.py` | Handle `queued` → `in-progress` transition | | `release_issue` | `server/tools/tools_issue_lifecycle.py` | Remove `queued` in cleanup paths | | `swap_labels` | `execution/github.py` | No change needed (generic) | | `process-issues` | `skills_extended/process-issues/SKILL.md` | Phase 0.5 uses `queued`, Step 3 swaps to `in-progress` | | `build-execution-map` | `skills_extended/build-execution-map/SKILL.md` | Query `queued` label alongside `in-progress` for conflict detection | ### R5: Classification vs. Lifecycle Label Distinction The system should distinguish between: - **Lifecycle labels** (`queued`, `in-progress`, `staged`, `fail`) — managed by the state machine, subject to atomic swaps and transition validation - **Classification labels** (`recipe:implementation`, `recipe:remediation`, `bug`, `enhancement`, `autoreported`) — applied by triage/reporting, never removed by the pipeline, not part of the state machine This distinction should be encoded in the architecture, not just documented. ## Implementation Plan Plan files: - `/home/talon/projects/autoskillit-runs/impl-20260508-023057-647423/.autoskillit/temp/make-plan/formalize_issue_label_lifecycle_plan_2026-05-08_023500_part_a.md` - `/home/talon/projects/autoskillit-runs/impl-20260508-023057-647423/.autoskillit/temp/make-plan/formalize_issue_label_lifecycle_plan_2026-05-08_023500_part_b.md` 🤖 Generated with [Claude Code](https://claude.com/claude-code) via AutoSkillit <!-- autoskillit:pipeline-signature steps=prepare_pr,run_arch_lenses,compose_pr,annotate_pr_diff,review_pr --> ## Token Usage Summary | Step | Model | count | uncached | output | cache_read | peak_ctx | turns | cache_write | time | |------|-------|-------|----------|--------|------------|----------|-------|-------------|------| | plan | claude-sonnet-4-6 | 1 | 80 | 32.2k | 840.0k | 81.9k | 100 | 82.2k | 18m 26s | | verify | claude-sonnet-4-6 | 2 | 1.9k | 27.0k | 1.9M | 77.8k | 210 | 108.7k | 15m 53s | | implement* | MiniMax-M2.7-highspeed | 2 | 5.8M | 44.2k | 4.9M | 87.5k | 374 | 272.0k | 25m 15s | | fix | claude-sonnet-4-6 | 1 | 394 | 21.5k | 3.3M | 98.0k | 126 | 98.3k | 17m 13s | | prepare_pr* | MiniMax-M2.7-highspeed | 1 | 91.2k | 8.3k | 206.6k | 34.2k | 25 | 60.5k | 2m 7s | | compose_pr* | MiniMax-M2.7-highspeed | 1 | 44.8k | 2.3k | 169.6k | 28.7k | 14 | 15.1k | 55s | | **Total** | | | 6.0M | 135.5k | 11.2M | 98.0k | | 636.9k | 1h 19m | \* *Step used a non-Anthropic provider; caching behavior may differ.* ## Token Efficiency | Step | LoC Changed | cache_read/LoC | cache_write/LoC | output/LoC | |------|-------------|----------------|-----------------|------------| | plan | 0 | — | — | — | | verify | 0 | — | — | — | | implement | 604 | 8054.5 | 450.4 | 73.2 | | fix | 52 | 62593.5 | 1891.2 | 413.6 | | prepare_pr | 0 | — | — | — | | compose_pr | 0 | — | — | — | | **Total** | **656** | 17128.3 | 970.9 | 206.5 | ## Model Usage Breakdown | Model | steps | uncached | output | cache_read | cache_write | time | |-------|-------|----------|--------|------------|-------------|------| | claude-sonnet-4-6 | 3 | 2.3k | 80.7k | 6.0M | 289.2k | 51m 34s | | MiniMax-M2.7-highspeed | 3 | 6.0M | 54.8k | 5.2M | 347.7k | 28m 17s | --------- Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
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.
Promotion: develop to main
This release represents a major expansion of AutoSkillit with 1,459 commits across 200 PRs, introducing three new core packages:
fleet/for campaign dispatch,workspace/for clone management, andplanner/for progressive resolution planning. Performance has been significantly improved through adoption of orjson, regex, uvloop, CDumper for YAML writes, CSafeLoader (11.1x speedup), and pre-compiled recipe YAML, collectively reducing hot-path latency across the board. The L2/L3 orchestration layers gained substantial new capabilities including activity-aware fleet dispatch timeouts, resumable quota sleep exits, L2 session resume for context exhaustion, and a full fleet campaign retry and preview system. Reliability was hardened through a dozen targeted Rectify fixes addressing resume gate boot failures, review loop counter bypasses, merge conflict semantic guards, and token summary schema drift. The recipe schema, hook structure, and CLI surface have all been meaningfully extended with new kinds, fields, validation commands, and the sous-chef → admiral rename for L3 terminology clarity.Stats: 1,550 files changed, +287,137 / -46,998 lines | 1,459 commits | 200 PRs | 179 fixes, 480 features, 27 refactors, 22 infra, 50 tests, 9 docs
Highlights
fleet/(campaign dispatch + semaphore + sidecar + liveness),workspace/(clone management + worktrees + skill resolution), andplanner/(progressive resolution planner with phases, assignments, and work packages)guards/(17 scripts) andformatters/(5 files); 9 new runtime dependencies addedRelease Notes
New Features
fleet/package with campaign dispatch, semaphore, sidecar, liveness probes, and state persistence; includes pre-launch preview and retry unblocking from terminal FAILURE stateworkspace/(clone management, worktrees, skill resolution) andplanner/(progressive resolution with phases, assignments, WPs, validation) added as first-class packagesautoskillit config validatecommand surfaces misconfiguration before runtimeBug Fixes
Performance
reon hot paths — PCRE2-backed with significant throughput gains on complex patternsRefactoring
Infrastructure
guards/(17 scripts) andformatters/(5 files) subdirectories;HOOK_REGISTRYandRETIRED_SCRIPT_BASENAMESupdated accordinglyintegration→developreviewconfig section — recipe-level review configuration surface added to AutomationConfigBreaking Changes
/autoskillit:sous-chefmust be updated to/autoskillit:admiralhooks/*.pynow live underhooks/guards/orhooks/formatters/; any external hook registrations must be updatedsafety.protected_branches— the branch nameintegrationhas been replaced bydevelop; update any config files that reference the old valueAttention Required
autoskillit initfrom within a git worktree; usetask install-worktreeinsteadCLAUDE_CODE_EXIT_AFTER_STOP_DELAY=120000— without this env var, subagents may not exit cleanly when finishedMerged PRs
Show all 200 PRs
The full list of 200 merged PRs includes additional features, fixes, tests, and infrastructure changes spanning v0.7.0 through v0.9.562. See the commit history for the complete changelog.
Linked Issues
160+ additional linked issues
This promotion carries forward closing references for 160+ additional issues across all domains. See the closing references section below.
Attention Required
validate_reciperenamed tovalidate_recipe_structureintegrationtodevelop; quota_guard dual-window fields addedArchitecture Impact
Module Dependency (Structural — "How are modules coupled?")
%%{init: {'flowchart': {'nodeSpacing': 50, 'rankSpacing': 70, 'curve': 'basis'}}}%% graph TB classDef cli fill:#1a237e,stroke:#7986cb,stroke-width:2px,color:#fff; classDef stateNode fill:#004d40,stroke:#4db6ac,stroke-width:2px,color:#fff; classDef handler fill:#e65100,stroke:#ffb74d,stroke-width:2px,color:#fff; classDef phase fill:#6a1b9a,stroke:#ba68c8,stroke-width:2px,color:#fff; classDef newComponent fill:#2e7d32,stroke:#81c784,stroke-width:2px,color:#fff; classDef output fill:#00695c,stroke:#4db6ac,stroke-width:2px,color:#fff; classDef detector fill:#b71c1c,stroke:#ef5350,stroke-width:2px,color:#fff; classDef integration fill:#c62828,stroke:#ef9a9a,stroke-width:2px,color:#fff; subgraph IL3 ["IL-3 — APPLICATION"] direction LR SERVER["● server/<br/>━━━━━━━━━━<br/>FastMCP server<br/>★ tools/ subpackage<br/>● _factory.py, _state.py"] CLI["● cli/<br/>━━━━━━━━━━<br/>CLI entry points<br/>★ doctor/, fleet/, session/, ui/<br/>★ 11 new modules"] end subgraph IL2 ["IL-2 — DOMAIN"] direction LR RECIPE["● recipe/<br/>━━━━━━━━━━<br/>Schema + validation<br/>★ rules/ subpackage<br/>● schema.py — RecipeKind, RecipeBlock<br/>Fan-in: 24"] MIGRATION["● migration/<br/>━━━━━━━━━━<br/>Versioned migration engine<br/>● engine.py — AdvisoryAdapter"] FLEET["★ fleet/<br/>━━━━━━━━━━<br/>Campaign dispatch<br/>semaphore, sidecar, liveness<br/>11 modules"] end subgraph IL1 ["IL-1 — INFRASTRUCTURE"] direction LR CONFIG["● config/<br/>━━━━━━━━━━<br/>AutomationConfig + Dynaconf<br/>● defaults.yaml<br/>Fan-in: 29"] PIPELINE["● pipeline/<br/>━━━━━━━━━━<br/>ToolContext DI, gate<br/>● tokens.py, telemetry_fmt.py"] EXECUTION["● execution/<br/>━━━━━━━━━━<br/>Headless sessions, CI<br/>★ process/, headless/<br/>★ session/, merge_queue/<br/>35 modules → core"] WORKSPACE["★ workspace/<br/>━━━━━━━━━━<br/>Clone mgmt, worktrees<br/>skill resolution<br/>9 modules"] PLANNER["★ planner/<br/>━━━━━━━━━━<br/>Progressive resolution<br/>phases, assignments, WPs<br/>5 modules"] end subgraph IL0 ["IL-0 — FOUNDATION"] direction LR CORE["● core/<br/>━━━━━━━━━━<br/>★ types/ subpackage<br/>★ runtime/ subpackage<br/>✕ _type_*.py replaced<br/>Fan-in: 200"] end subgraph HOOKS_LAYER ["HOOKS — Cross-cutting"] direction LR HOOKS["● hooks/<br/>━━━━━━━━━━<br/>★ guards/ — 17 scripts<br/>★ formatters/ — 5 files<br/>★ _dispatch, _hook_settings"] HOOKREG["● hook_registry.py<br/>━━━━━━━━━━<br/>Registry + retired names"] end SERVER -->|"imports"| RECIPE SERVER -->|"imports"| MIGRATION CLI -->|"imports"| RECIPE SERVER -->|"imports"| CONFIG SERVER -->|"imports"| PIPELINE SERVER -->|"imports"| EXECUTION SERVER -->|"imports"| WORKSPACE CLI -->|"imports"| CONFIG CLI -->|"imports"| EXECUTION CLI -->|"imports"| WORKSPACE SERVER -->|"imports"| HOOKS CLI -->|"imports"| HOOKS CLI -->|"imports"| HOOKREG RECIPE -->|"imports"| CORE MIGRATION -->|"imports"| CORE FLEET -->|"imports"| CORE FLEET -.->|"lateral: _prompts"| HOOKS CONFIG -->|"imports"| CORE PIPELINE -->|"imports"| CORE EXECUTION -->|"imports"| CORE WORKSPACE -->|"imports"| CORE PLANNER -->|"imports"| CORE PIPELINE -.->|"runtime: AutomationConfig"| CONFIG HOOKS -->|"imports"| CORE HOOKREG -->|"imports"| CORE SERVER -->|"imports"| CORE CLI -->|"imports"| CORE class SERVER,CLI cli; class RECIPE,MIGRATION phase; class FLEET newComponent; class CONFIG,PIPELINE handler; class EXECUTION handler; class WORKSPACE,PLANNER newComponent; class CORE stateNode; class HOOKS,HOOKREG detector;Process Flow (Physiological — "How does it behave?")
%%{init: {'flowchart': {'nodeSpacing': 40, 'rankSpacing': 55, 'curve': 'basis'}}}%% flowchart TB classDef terminal fill:#1a237e,stroke:#7986cb,stroke-width:2px,color:#fff; classDef stateNode fill:#004d40,stroke:#4db6ac,stroke-width:2px,color:#fff; classDef handler fill:#e65100,stroke:#ffb74d,stroke-width:2px,color:#fff; classDef phase fill:#6a1b9a,stroke:#ba68c8,stroke-width:2px,color:#fff; classDef newComponent fill:#2e7d32,stroke:#81c784,stroke-width:2px,color:#fff; classDef detector fill:#b71c1c,stroke:#ef5350,stroke-width:2px,color:#fff; classDef output fill:#00695c,stroke:#4db6ac,stroke-width:2px,color:#fff; START([MCP Tool Call]) COMPLETE([SkillResult JSON]) ERROR([Gate Error / Crash]) subgraph Visibility ["● Session-Type Tag Visibility"] direction TB STYPE{"Session Type?<br/>━━━━━━━━━━<br/>env: SESSION_TYPE"} FLEET_VIS["★ Fleet Tags<br/>━━━━━━━━━━<br/>fleet-dispatch"] ORCH_VIS["Orchestrator Tags<br/>━━━━━━━━━━<br/>kitchen-core + packs"] SKILL_VIS["Skill Tags<br/>━━━━━━━━━━<br/>headless only"] end subgraph GatePhase ["● Gate Lifecycle"] direction TB OPEN_K["● open_kitchen<br/>━━━━━━━━━━<br/>gate.enable + hook_config<br/>quota cache prime"] GATE{"Gate<br/>Enabled?"} LOAD_R["● load_and_validate<br/>━━━━━━━━━━<br/>Cache → YAML → sub-recipe<br/>→ semantic rules → contract"] VALID{"Recipe<br/>Valid?"} end subgraph Dispatch ["● run_skill Dispatch"] direction TB GUARDS["● Guard Chain<br/>━━━━━━━━━━<br/>1. orchestrator_check<br/>2. gate_enabled<br/>3. skill_command_valid<br/>4. cwd_absolute<br/>5. dry_walkthrough"] PROVIDER["● Provider Resolution<br/>━━━━━━━━━━<br/>step override → recipe<br/>→ YAML field → config"] SKILL_SETUP["● Skill Session Setup<br/>━━━━━━━━━━<br/>resolve namespace<br/>compute closure<br/>init ephemeral dir"] end subgraph Headless ["● Headless Session"] direction TB LAUNCH["● _execute_claude_headless<br/>━━━━━━━━━━<br/>subprocess + PTY<br/>completion marker"] STALE_CHK{"Termination<br/>Reason?"} RECOVER["● Recovery Paths<br/>━━━━━━━━━━<br/>Channel B drain-race<br/>marker search<br/>pattern recovery"] end subgraph Adjudication ["● Result Adjudication"] direction TB OUTCOME["● _compute_outcome<br/>━━━━━━━━━━<br/>success gate chain<br/>+ retry FSM"] RETRY{"needs_retry?"} BUDGET{"Budget<br/>Exhausted?"} FALLBACK["● Provider Fallback<br/>━━━━━━━━━━<br/>inject fallback env<br/>re-launch session"] POST["● Post-Session<br/>━━━━━━━━━━<br/>flush log, record tokens<br/>refresh quota cache"] end START --> STYPE STYPE -->|"FLEET"| FLEET_VIS STYPE -->|"ORCHESTRATOR"| ORCH_VIS STYPE -->|"SKILL"| SKILL_VIS FLEET_VIS --> OPEN_K ORCH_VIS --> OPEN_K SKILL_VIS --> OPEN_K OPEN_K --> GATE GATE -->|"closed"| ERROR GATE -->|"open"| LOAD_R LOAD_R --> VALID VALID -->|"errors"| ERROR VALID -->|"valid"| GUARDS GUARDS -->|"any guard fails"| ERROR GUARDS -->|"pass"| PROVIDER PROVIDER --> SKILL_SETUP SKILL_SETUP --> LAUNCH LAUNCH --> STALE_CHK STALE_CHK -->|"STALE / IDLE_STALL"| RECOVER STALE_CHK -->|"TIMED_OUT"| OUTCOME STALE_CHK -->|"COMPLETED / NATURAL_EXIT"| OUTCOME RECOVER --> OUTCOME OUTCOME --> RETRY RETRY -->|"no"| POST RETRY -->|"yes"| BUDGET BUDGET -->|"exhausted"| POST BUDGET -->|"remaining"| FALLBACK FALLBACK --> LAUNCH POST --> COMPLETE class START,COMPLETE,ERROR terminal; class STYPE,GATE,VALID,STALE_CHK,RETRY,BUDGET stateNode; class OPEN_K,LOAD_R,GUARDS,PROVIDER,SKILL_SETUP phase; class LAUNCH,RECOVER,OUTCOME handler; class FLEET_VIS,ORCH_VIS,SKILL_VIS newComponent; class FALLBACK,POST output;Closes #1622
Closes #1695
Closes #1699
Closes #1700
Closes #1701
Closes #1702
Closes #1703
Closes #1706
Closes #1707
Closes #1708
Closes #1709
Closes #1710
Closes #1712
Closes #1716
Closes #1717
Closes #1718
Closes #1719
Closes #1722
Closes #1723
Closes #1724
Closes #1725
Closes #1726
Closes #1727
Closes #1728
Closes #1729
Closes #1735
Closes #1745
Closes #1747
Closes #1748
Closes #1749
Closes #1751
Closes #1752
Closes #1753
Closes #1754
Closes #1755
Closes #1756
Closes #1772
Closes #1773
Closes #1774
Closes #1775
Closes #1776
Closes #1777
Closes #1778
Closes #1779
Closes #1780
Closes #1798
Closes #1802
Closes #1803
Closes #1804
Closes #1805
Closes #1806
Closes #1825
Closes #1831
Closes #1834
Closes #1835
Closes #1837
Closes #1838
Closes #1849
Closes #1851
Closes #1852
Closes #1853
Closes #1860
Closes #1861
Closes #1862
Closes #1863
Closes #1875
Closes #1877
Closes #1879
Closes #1880
Closes #1881
Closes #1882
Closes #1883
Closes #1884
Closes #1885
Closes #1886
Closes #1887
Closes #1888
Closes #1897
Closes #1898
Closes #1899
Closes #1900
Closes #1901
Closes #1902
Closes #1903
Closes #1905
Closes #1906
Closes #1910
Closes #1918
Closes #1924
Closes #1928
Closes #1932
Closes #1936
Closes #1943
Closes #1944
Closes #1945
Closes #1954
Closes #1955
Closes #1963
Closes #1964
Closes #1965
Closes #1966
Closes #1975
Closes #1976
Closes #1980
Closes #1986
Closes #1987
Closes #2005
Closes #2007
Closes #2008
Closes #2009
Closes #2020
Closes #2029
Closes #2035
Closes #2036
Closes #2039
Closes #2043
Closes #2044
Closes #2045
Closes #2047
Closes #2048
Closes #2049
Closes #2051
Closes #2061
Closes #2063
Closes #2097
Closes #2133
Closes #2134
Closes #2136
Closes #2137
Closes #2138
Closes #2139
Closes #2140
Closes #2141
Closes #2158
Closes #2173
Closes #2182
Closes #2188
Closes #2190
Closes #2196
Closes #2197
Closes #2205
Closes #2208
Closes #720
Closes #830
Closes #831
Closes #832
Closes #833
Closes #834
Closes #835
Closes #836
Closes #837
Closes #838
Closes #839
Closes #840
Closes #841
Closes #842
Closes #845
Closes #847
Closes #848
Closes #849
Closes #850
Closes #851
Closes #852
Closes #857
Closes #858
Generated with Claude Code via AutoSkillit