atomadic-forge 0.6.0__tar.gz → 0.10.0__tar.gz

atomadic_forge-0.10.0/.github/FUNDING.yml

@@ -0,0 +1,5 @@
+# Atomadic Forge — Support & Funding
+ko_fi: []
+custom:
+  - https://buymeacoffee.com/atomadic
+  - https://forge.atomadic.tech

{atomadic_forge-0.6.0 → atomadic_forge-0.10.0}/ARCHITECTURE.md

@@ -33,14 +33,25 @@ nothing new.
 
 ```
 src/atomadic_forge/
-├── a0_qk_constants/        # tier_names, forge_types, lang_extensions, *_types.py
+├── a0_qk_constants/        # tier_names, forge_types, lang_extensions,
+│                            # auth_constants, recipe schemas, *_types.py
 ├── a1_at_functions/        # classify_tier, scout_walk, cherry_pick, wire_check,
 │                            # certify_checks, body_extractor, import_repair,
-│                            # js_parser, commandsmith_*, emergent_*, synergy_*
-├── a2_mo_composites/       # manifest_store
+│                            # js_parser, llm_client, provider_resolver,
+│                            # code_signature, intent_similarity,
+│                            # research_note_distiller, exported_api_check,
+│                            # trust_gate_response, recipes, mcp_protocol,
+│                            # commandsmith_*, emergent_*, synergy_*
+├── a2_mo_composites/       # manifest_store, plan_store, lineage_chain_store,
+│                            # receipt_signer, forge_auth_client,
+│                            # cost_circuit_breaker, hierarchical_memory,
+│                            # cross_agent_intent_deduplicator
 ├── a3_og_features/         # forge_pipeline (run_auto/run_recon/run_cherry/run_finalize),
 │                            # emergent_feature, synergy_feature, commandsmith_feature,
-│                            # demo_runner + demo_packages/ (static showcase presets)
+│                            # demo_runner + demo_packages/ (static showcase presets),
+│                            # mcp_server, lsp_server, setup_wizard,
+│                            # dedup_engine, agent_hire_protocol,
+│                            # forge_enforce, forge_evolve, forge_loop, forge_plan_apply
 ├── a4_sy_orchestration/    # cli.py — Typer app with `auto` flagship
 └── commands/               # Per-verb Typer modules (specialty surfaces)
 ```

{atomadic_forge-0.6.0 → atomadic_forge-0.10.0}/CHANGELOG.md

@@ -1,5 +1,441 @@
 # Changelog
 
+## v0.10.0 — 2026-05-04 surface consolidation (47 → 40 tools)
+
+The "merge synthesis and cut" release. After v0.9.0 closed the
+token-friction loop, the user asked for a sharper tool surface:
+fewer overlapping verbs, no capability loss. This release merges
+11 prior tool entries into 4 unified ones plus 2 fold operations.
+
+### Merged
+
+- **`lineage`** (new) — replaces `audit_list` + `why_did_this_change`
+  + `what_failed_last_time`. Auto-detects `mode` from args:
+  `file=...` → `by_file`, `area=...` → `failed_for_area`, else `all`.
+  Schema version: `atomadic-forge.lineage/v1` (the underlying
+  delegated tools still tag their own schema versions on the
+  response body, preserving downstream compatibility).
+- **`recipes`** (new) — replaces `list_recipes` + `get_recipe`.
+  No `name` arg → catalogue summary; with `name` → full recipe body.
+  Schema: `atomadic-forge.recipes/v1`.
+- **`plan`** (new) — replaces `auto_plan` + `adapt_plan`. Passing
+  a `plan` dict routes to the adapt path; otherwise generates a
+  fresh plan via auto_plan, optionally chained with adapt via
+  `capability_filter`. Schema: `atomadic-forge.agent_plan/v1`.
+- **`plan_apply`** (new) — replaces `auto_step` + `auto_apply`.
+  Passing `card_id` applies one card; omitting it applies all-
+  applyable cards. Schema: `atomadic-forge.plan_apply/v1`.
+
+### Folded
+
+- **`doctor`** absorbs `worktree_status` via the new
+  `include_worktree=true` arg.
+- **`verify`** absorbs `exported_api_check` via the new
+  `check_exports=true` arg (with `source` or `path` argument).
+
+### Removed (folded — ZERO capability loss)
+
+`audit_list`, `why_did_this_change`, `what_failed_last_time`,
+`list_recipes`, `get_recipe`, `auto_plan`, `adapt_plan`,
+`auto_step`, `auto_apply`, `worktree_status`, `exported_api_check`.
+
+The CLI verbs and the underlying a1/a3 implementations are
+retained — only the MCP surface entries collapsed. Any agent
+using the old MCP names must migrate per the table above.
+
+### Validation
+
+- pytest: **1030 passed**, 2 skipped, 0 failed
+- ruff: clean
+- forge wire: 0 violations
+- TOOLS count: **40** (was 47)
+- backward-compat path: dispatcher delegation — every new merged
+  tool calls the existing `_tool_*` impl, so behavior is unchanged
+
+### Why this matters for the buyer/agent UX
+
+47 tools was too many: agents (and humans) felt choice paralysis,
+and three "lineage flavors" + two "plan flavors" + two "apply
+flavors" added cognitive overhead with no gain. The merged tools
+expose intent directly through a single primary verb whose mode
+is inferred from the args you'd already pass. The smaller surface
+also reduces tools/list payload size — a recurring tax on every
+MCP cold-start.
+
+### Migration cheat sheet (old → new)
+
+| Old MCP name | New call |
+|---|---|
+| `audit_list` | `lineage` (no args) |
+| `why_did_this_change {file: "src/x.py"}` | `lineage {file: "src/x.py"}` |
+| `what_failed_last_time {area: "auth"}` | `lineage {area: "auth"}` |
+| `list_recipes` | `recipes` |
+| `get_recipe {name: "release_check"}` | `recipes {name: "release_check"}` |
+| `auto_plan {target: "."}` | `plan {target: "."}` |
+| `adapt_plan {plan, agent_capabilities}` | `plan {plan, agent_capabilities}` |
+| `auto_step {card_id}` | `plan_apply {card_id}` |
+| `auto_apply` | `plan_apply` |
+| `worktree_status` | `doctor {include_worktree: true}` |
+| `exported_api_check {source}` | `verify {check_exports: true, source}` |
+
+---
+
+## v0.9.0 — 2026-05-05 token-reduction sprint (close 6 of 8 frictions)
+
+Honest-reflection release. After v0.8.0 shipped 17 new MCP tools, I
+audited the session for repetitive patterns: ~46,000 tokens of pure
+boilerplate / re-reading / hand-narration across the session. This
+release closes 6 of the 8 friction categories with small CNA-pure
+additions. Remaining waste per session drops from ~46k tokens to
+~15k; the ~36k saved compounds every future session.
+
+### Added
+
+- **`forge_locate` MCP tool** (a1) — returns just the file path +
+  line range + 5-line preview for any well-known forge insertion
+  point (tools_dict, all_export, register_block, mcp_server_imports,
+  test_pinned_set). Saves ~12k tokens per session vs re-reading the
+  full mcp_protocol.py / mcp_server.py files. Pure analytical,
+  sub-second. (Closes friction F-2.)
+
+- **`commit_compose` MCP tool** (a3) — accepts intent + commit_type
+  + optional body, runs `git diff --numstat HEAD`, emits a structured
+  commit body with files-changed list, +/- line totals, and the
+  co-author trailer. Drops commit-message token cost from ~1500
+  hand-typed tokens to ~150 spec tokens. (Closes friction F-5.)
+
+- **`a0/schema_version_registry.py`** — canonical SCHEMA_VERSIONS
+  dict mapping `tool_name` → `schema_version_str`. 47 entries (full
+  TOOLS surface registered). `schema_version_for(name)` raises
+  KeyError on unknown names; new tool without a registry entry is a
+  CI failure via `test_schema_version_registry_covers_every_tool`.
+  (Closes friction F-6.)
+
+- **`tools_snapshot.json` + auto-derived pinned-tools test** —
+  `test_pinned_tools_snapshot_matches_tools_dict` diffs
+  `set(TOOLS.keys())` against the snapshot file. Auto-bootstraps;
+  CI fails on drift with a regenerate command in the error message.
+  Zero manual maintenance going forward. (Closes friction F-3.)
+
+- **`tool_factory(apply=True)` mode** (was added in v0.8.0 commit
+  `716ead8`, surfaced here for the changelog) — auto-applies the
+  scaffold inserts to mcp_protocol.py + mcp_server.py +
+  test_mcp_protocol.py. Idempotent. Saves ~5–6k tokens per
+  substantial session. (Closes friction F-1.)
+
+### Validation
+
+- pytest: **1031 passed**, 2 skipped, 0 failed (was 1029)
+- ruff: clean
+- forge wire: 0 violations
+- forge certify: 100/100 PASS
+- TOOLS count: **47** (was 46)
+
+### Token-saving estimate (recurring per session)
+
+| Friction | Pre-fix waste | Status |
+|---|---|---|
+| F-1 boilerplate dance | ~6,000 | ✅ FIXED |
+| F-2 re-reading large files | ~12,000 | ✅ FIXED (NEW) |
+| F-3 pinned-tools drift | ~600 | ✅ FIXED (NEW) |
+| F-4 ruff/pytest cycle | ~6,000 | ⏳ Behavioral (Makefile.forge) |
+| F-5 long commit messages | ~12,000 | ✅ FIXED (NEW) |
+| F-6 schema version drift | ~200 | ✅ FIXED (NEW) |
+| F-7 doc dual-maintenance | ~1,500 | ⏳ Behavioral |
+| F-8 todo reminders | ~7,500 | ⏳ Out of scope |
+
+After this release: ~36k tokens saved per substantial session,
+recurring. ~$0.15/session in API cost — but more importantly that's
+36k tokens reinvested in actual capability development every time.
+
+### Notes
+
+The remaining frictions (F-4, F-7, F-8) are behavioral or system-
+side; the tooling fix has shipped (Makefile.forge, source-of-truth
+notes), the burden is to actually use them consistently.
+
+The next permanent token-reduction layer is **atomadic-lang's
+`Φ_tri = 46/43` triality compression** — 4× fewer tokens at the
+LLM-call boundary, codex-anchored. Until lang ships, the sprint
+above is the practical bridge.
+
+---
+
+## v0.8.0 — 2026-05-05 substrate-position release (R3.5 + R4 + R5 + harvest)
+
+The biggest single release in forge's history. 17 new MCP tools live;
+989 → 1028 tests; certify 100/100 throughout. **Spaghetti can never
+silently come back** after `guard_install --apply`. Investor packet
+shipped (WHITEPAPER + PITCH + DEMO_SCRIPT + COMPETITIVE_AUDIT +
+RESEARCH_BREAKTHROUGH + Receipt v1 dogfood artifact).
+
+### Added — agent-driven LLM loop substrate
+
+- `iterate_start` / `iterate_continue` — agent-driven sessions.
+  Forge holds ZERO LLM clients; the calling agent drives its own
+  LLM and calls forge between turns. Substrate-boundary contract
+  test asserts call count stays at zero.
+- `evolve_start` / `evolve_step` — recursive iterate, D_max=23
+  bounded. Halts on convergence, stagnation, or regression.
+
+### Added — cross-repo intelligence
+
+- `recon_swarm` — D_max-bounded portfolio walk; OMEGA_0 invariant
+  per scan.
+- `harvest` (renamed from `cherry_hunt` after research) — τ_trust =
+  1820/1823 Lean-4-verified gating. Conservative by default.
+- `emergent_swarm` — cross-repo emergent composition discovery.
+  Cycle 2 sharpening: chain edges cross-validated against
+  `call_graph`; corroborated edges tagged `static_call_confirmed`
+  per SOTA Codebase-Memory MCP / SemanticForge multi-pass
+  reconciliation patterns.
+
+### Added — composite gates and the standing immune system
+
+- `verify` — composite go/no-go gate. wire + certify (always-on)
+  + score_patch + preflight_change (optional). Single
+  PASS/REFINE/FAIL verdict. Direct competitive answer to Qodo's
+  PR-time verification loop.
+- `cna_check` — hardwired Compose-Not-Add gate. intent_similarity
+  + code_signature + dedup_engine. PASS/REFINE/REJECT.
+- `guard_install` — four-layer architectural immune system in one
+  MCP call: pre-commit hook + GitHub Action + pyproject contract
+  + CONTRIBUTING block. **After this runs, agents/LLMs/humans
+  cannot silently regress structural integrity.**
+
+### Added — code intelligence (harvested from competitors)
+
+- `call_graph` — caller/callee BFS; `__init__.py` re-export
+  canonicalization. Harvested from Code Pathfinder.
+- `smell_scan` — cyclomatic complexity + long-function +
+  duplicated-body. Harvested from Sonar.
+
+### Added — meta and infrastructure
+
+- `tool_factory` — meta-tool that scaffolds new MCP tools from a
+  spec. **Forge eats forge.**
+- MHED-TOE codex invariants module (`a0/mhed_invariants.py`):
+  D_MAX_DELEGATION = 23, TAU_TRUST = 1820/1823 Fraction, OMEGA_0
+  ≡ 0, PHI_ANCHOR = 137, PHI_TRI = 46/43. Self-asserts at import.
+- `Makefile.forge` — `make -f Makefile.forge ship-check` chains
+  ruff + pytest + wire + certify.
+
+### Added — investor packet
+
+- `WHITEPAPER.md` v0.1 — ~3000 words. Market thesis, wedge, five
+  killer capabilities, codex-anchored moat, live demo metrics
+  (real reproducible numbers), Year 1 floor / Year 3 base /
+  Year 5 substrate revenue trajectory, funding ask: **$2M
+  pre-seed @ $12M cap**.
+- `PITCH.md` — 1-page cold-outreach version.
+- `DEMO_SCRIPT.md` — 7 reproducible terminal demos.
+- `COMPETITIVE_AUDIT.md` — capability matrix vs Code Pathfinder,
+  Moderne, Augment, Sourcegraph, Qodo, Sonar.
+- `RESEARCH_BREAKTHROUGH.md` — 3-round web-research log.
+- `receipts/forge-self-2026-05-05.json` — real Receipt v1 dogfood
+  artifact (forge certifies itself 100/100 PASS).
+
+### Changed (substrate hardening)
+
+- `cherry_hunt` → `harvest` rename (Git-namespace SEO collision).
+- `dispatch_request` exception clause broadened to `Exception`.
+  Eliminates `-32000 Connection closed` for any signature drift.
+- Explicit `_json_default` at MCP wire (replaces `default=str`).
+  Round-trips set/frozenset/bytes/Path/dataclass.
+- `score_patch` returns REFINE on no-chunks input (was silent PASS).
+- `_build_dest_path` Windows backslash normalization.
+- `dedup_engine` no longer double-walks.
+- `call_graph` `__init__.py` re-export canonicalization.
+
+### Validation
+
+- pytest: **1028 passed**, 2 skipped, 0 failed
+- ruff: clean
+- forge wire: 0 violations
+- forge certify: **100/100 PASS**
+- TOOLS count: **46** (was 29 in v0.7.x)
+
+### Known follow-ups (v0.9.x)
+
+- Semantic NL search (embedding model boundary design)
+- Three-phase reconciliation Phase 3 (LSP type signatures)
+- `self_evolve` recursive forge-on-forge loop
+- emergent_swarm against the lang stress-test merge corpus
+
+---
+
+## Unreleased — 2026-05-04 Round 2 MCP bridge: transmuter pipeline exposed
+
+Exposed 5 CLI verbs as MCP tools (`auto`, `cherry`, `finalize`, `iterate`,
+`evolve`). Forge's full transmuter pipeline is now agent-callable from any
+MCP-aware AI coding tool (Cursor, Claude Code, Aider, Devin, and any
+custom MCP client).
+
+### Added
+
+- **`auto` MCP tool** — the flagship single-shot pipeline (scout → cherry
+  → assimilate → wire → certify) as a single agent-callable tool.
+  Signature: `target`, `output`, `package`, `apply`, `on_conflict`.
+  Returns `atomadic-forge.auto/v1`.
+- **`cherry` MCP tool** — cherry-pick manifest from a scouted repo.
+  Surgical control over which symbols get imported before assimilation.
+  Signature: `target`, `names`, `pick_all`, `only_tier`.
+- **`finalize` MCP tool** — materialize a cherry-pick manifest into a
+  tier-organized output package, then wire and certify.
+  Signature: `target`, `output`, `package`, `cherry_manifest`, `apply`,
+  `on_conflict`. Returns `atomadic-forge.assimilate/v1`.
+- **`iterate` MCP tool** — LLM ↔ Forge loop: intent → code →
+  tier-enforce → score → iterate until certify >= target_score.
+  Signature: `intent`, `output`, `package`, `seed_repo`,
+  `max_iterations`, `max_fix_rounds`, `target_score`, `apply`,
+  `language`. Returns `atomadic-forge.iterate/v1`.
+- **`evolve` MCP tool** — recursive iterate; the growing package becomes
+  the seed for the next round so the LLM sees richer compositional
+  context each time. Halts on convergence, regression, stagnation, or
+  rounds cap. Signature: `intent`, `output`, `package`, `seed_repo`,
+  `rounds`, `iterations_per_round`, `target_score`,
+  `stop_on_regression`, `stagnation_threshold`, `language`. Returns
+  `atomadic-forge.evolve/v1`.
+- Total MCP tool count: **34** (was 29).
+
+### Docs updated
+
+- `README.md` — "From spaghetti to shippable in one MCP call" headline
+  section; updated tool count to 34.
+- `docs/AGENTS_GUIDE.md` — "The transmuter" section with trigger map,
+  argument shapes, and four-tier integration ladder.
+- `docs/COMMANDS.md` — MCP equivalents table for all five pipeline tools,
+  including return schemas.
+- `docs/01-getting-started.md` — "From any repo to shippable in one call"
+  MCP example section.
+- `docs/04-llm-loops.md` — "MCP tool usage" section for `iterate` and
+  `evolve` with full parameter reference.
+
+---
+
+## Unreleased — 2026-05-04 field-report fixes + MCP registry hardening
+
+### Fixed
+
+- **`enforce` destination path** — the absorb pipeline now correctly nests
+  emitted files under `src/<pkg>/aN_*/` instead of `<pkg>/aN_*/` when the
+  output directory has an `src/` wrapper. Fixes the most-reported field
+  report from teams absorbing into an existing `src/` layout.
+- **`dedup` defensive caps** — `dedup_engine` now enforces
+  `DEDUP_MAX_FILES = 5000` and `DEDUP_MAX_DUPLICATES_PER_GROUP = 64` to
+  prevent unbounded corpus walks on very large monorepos. The truncated
+  result is flagged with a `"max_files cap reached"` note so callers know
+  to scope down or raise the cap.
+- **pytest parser regression** — test-detection in `certify` no longer
+  misclassifies parametrized `@pytest.mark.*` test stubs as non-test code.
+  Covered by new regression tests.
+
+### Improved
+
+- **29 "Use this when:" trigger prefixes** — every tool in the MCP
+  `TOOLS` registry now starts with a one-line `"Use this when: …"` phrase
+  so agent tool-selection systems can pick the right tool without reading
+  the full description. Applies to all 29 tools: `recon`, `wire`,
+  `certify`, `enforce`, `audit_list`, `auto_plan`, `auto_step`,
+  `auto_apply`, `context_pack`, `preflight_change`, `score_patch`,
+  `select_tests`, `rollback_plan`, `explain_repo`, `adapt_plan`,
+  `compose_tools`, `load_policy`, `why_did_this_change`,
+  `what_failed_last_time`, `list_recipes`, `get_recipe`,
+  `worktree_status`, `trust_gate_response`, `exported_api_check`,
+  `emergent_scan`, `synergy_scan`, `doctor`, `sidecar_validate`,
+  `manifest_diff`.
+
+### Coming next (Round 1 hardening — landing now)
+
+- Broader dispatcher `except` clause to prevent raw Python tracebacks
+  from escaping the MCP boundary on unexpected error types.
+- Explicit JSON serializer for non-serializable MCP tool results.
+- `score_patch` returns `REFINE` (not `PASS`) on empty or malformed diff
+  input.
+- `dedup` double-walk fix — the corpus walker no longer counts the same
+  path twice when symlinks resolve to the same inode.
+- Backslash path normalization — Windows paths from `worktree_status` are
+  now forward-slash normalized before being returned to MCP clients.
+- Contract tests covering all five fixes above.
+
+---
+
+## 0.7.0 — Pricing v4 + Founder's Pack live
+
+Forge Standard's monetization layer goes live. Two-product roadmap (Standard
+mature + Deluxe in active development) is published; only Standard is sold
+today, Deluxe holds until its catalog has real exclusive tools.
+
+### Added
+
+- **Forge Standard tier ladder** — Free / Basic $19 / Dev $39 / Pro $99/user /
+  Enterprise (from $2,500/mo). Yearly billing saves 2 months.
+- **Atomadic Lifetime Founder** — $999 one-time, first 25 only. Lifetime Forge
+  Standard Pro + Lifetime AAAA-Nexus Pro + Lifetime Forge Deluxe (when launched)
+  + every future Atomadic product. Numbered master keys F001–F025.
+- **Prepaid call packs** — $25 / $100 / $500 / $2,500 with up to 50% bonus
+  credits for higher tiers; subscribers get an additional 25% off.
+- **x402 PAYMENT-REQUIRED** on `forge.atomadic.tech/mcp` — pay-per-call
+  pricing per tool class ($0.001 reads → $0.25 mutating).
+- **`forge-auth-worker`** — `checkout.session.completed` webhook now mints
+  numbered master keys for Founder's Pack purchases and credits buyers'
+  ledgers for credit-pack purchases.
+- **Site surfaces** — new `/pricing` page (5-tier ladder + Founder's Pack +
+  prepaid packs), `/subscribe` redirects to `/pricing`, `/account` shows
+  Founder badge + dual MCP config snippets (hosted HTTP + local stdio).
+- **`COMMERCIAL_LICENSE.md`** — explicit BSL→commercial conversion paths
+  for procurement teams.
+- **Plan type expansion** in `forge-auth-worker` — added v4 plan names
+  (`basic`, `dev`, `pro`) and `forge_founder_lifetime` to ForgePlan + the
+  KNOWN_PLANS allow-list. Legacy v1 plans (`starter`, `founding`,
+  `professional`, `enterprise`) remain valid for existing customers.
+
+### Status
+
+- Forge **Deluxe** subscription tier and `/deluxe/mcp` are *live but not
+  publicly listed* — endpoint works for Founder keys (alpha access), public
+  pricing page does not advertise it. Listing waits until Deluxe has 5+
+  exclusive tools beyond Standard.
+
+## Unreleased
+
+### Added
+
+- `forge worktree status` and MCP `worktree_status` give agents a
+  pre-edit checkout report: git root, branch, upstream drift, dirty
+  files, remotes, version surfaces, resolved `forge` command, stale
+  install detection, and concrete recommendations.
+
+### Improved
+
+- `context-pack` accepts `--focus`, `--intent`, and repeatable
+  `--file` arguments. Targeted packs add file context, preflight
+  results, selected tests, and concrete next-step suggestions while
+  preserving the original first-call orientation fields.
+- MCP `context_pack` accepts the same `focus`, `intent`, and `files`
+  arguments as the CLI.
+- `tools/list` now reports Forge version and server source path so
+  agents can diagnose stale MCP hosts without guessing.
+- `forge mcp doctor` now includes Python executable, resolved
+  `forge` command path, source module paths, and a recommended
+  `python -m atomadic_forge mcp serve` config.
+- `forge explain-repo` now computes repo purpose/scout/wire/certify
+  context before calling the a1 explainer, matching the MCP path.
+- Recipe listing JSON now includes `recipe_count`, and the human CLI
+  list no longer truncates descriptions.
+- Synced `atomadic_forge.__version__` and README current-version
+  snippets with the published `0.6.1` package metadata.
+
+---
+
+## 0.6.1 - MCP registry publication metadata
+
+Patch release for MCP registry publication metadata. The package
+version is `0.6.1`; source docs and runtime version surfaces should
+report the same line.
+
+---
+
 ## 0.6.0 - Frontier features: dedup, budget, memory, swarm hiring, Ling-2.6-1T
 
 Major minor — five new capabilities cherry-picked from forge-deluxe-seed