roam-code 13.5__tar.gz → 13.5.1__tar.gz

{roam_code-13.5/src/roam_code.egg-info → roam_code-13.5.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: roam-code
-Version: 13.5
+Version: 13.5.1
 Summary: Local codebase intelligence for AI coding agents: structural repo map, change-safety gates, MCP security receipts, and tamper-evident ChangeEvidence packets. 100% local, zero API keys, zero network egress.
 Author: Cranot
 License-Expression: Apache-2.0

{roam_code-13.5 → roam_code-13.5.1}/pyproject.toml

@@ -8,7 +8,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "roam-code"
-version = "13.5"
+version = "13.5.1"
 description = "Local codebase intelligence for AI coding agents: structural repo map, change-safety gates, MCP security receipts, and tamper-evident ChangeEvidence packets. 100% local, zero API keys, zero network egress."
 readme = "README.md"
 requires-python = ">=3.10"

{roam_code-13.5 → roam_code-13.5.1}/src/roam/commands/cmd_compile.py

@@ -11,7 +11,7 @@ the agent can consume.
 
 Empirically validated 2026-05-28: FactsEnvelope strictly dominates vanilla
 on capable models (Opus 4.8) — 99% of vanilla's quality at 54% of
-vanilla's cost. See internal/planning/FINAL-EVALUATION-2026-05-28.md.
+vanilla's cost.
 
 Output formats: ``--json`` (default), text.
 

{roam_code-13.5 → roam_code-13.5.1}/src/roam/commands/cmd_compile_stats.py

@@ -224,7 +224,7 @@ def _top_cache_misses(rows: list[dict], limit: int = 10) -> list[dict]:
     "--by-mode",
     is_flag=True,
     default=False,
-    help="W5 (2026-06-02) — group rows by agent_mode "
+    help="W5 — group rows by agent_mode "
     "(compile/roam/vanilla/unknown). Rows that pre-date the "
     "agent_mode field count as 'unknown'.",
 )
@@ -273,7 +273,7 @@ def compile_stats(
     # (one-off; not worth caching).
     if top_misses and rows:
         summary["top_cache_misses"] = _top_cache_misses(rows)
-    # W5 (2026-06-02) — by-mode breakdown. Joins on the `agent_mode` field
+    # W5 — by-mode breakdown. Joins on the `agent_mode` field
     # added to telemetry rows when ROAM_AGENT_MODE env var is set at compile
     # time (the host platform sets this per call). Pre-W5 rows lack the field,
     # so they bucket as 'unknown' — useful baseline.

{roam_code-13.5 → roam_code-13.5.1}/src/roam/commands/cmd_compiler_corpus.py

@@ -112,7 +112,7 @@ def _compile_one(prompt: str, cwd: str | None) -> dict:
 
     # "Probe empty" heuristic — delegated to the canonical introspector so
     # this command can never re-drift into the agent_contract.facts vs
-    # prefetched_facts bug (2026-06-02 dogfood). An L1 envelope is empty when
+    # prefetched_facts bug (dogfood). An L1 envelope is empty when
     # it carries no SUBSTANTIVE probe families; non-L1 labels are never
     # misses (lower-signal by design).
     from roam.plan.envelope_introspect import introspect as _introspect

{roam_code-13.5 → roam_code-13.5.1}/src/roam/mcp-server-card.json

@@ -2,7 +2,7 @@
   "$schema": "https://schemas.modelcontextprotocol.io/server-card/v1",
   "name": "roam-code",
   "title": "roam — instant codebase intelligence",
-  "version": "13.5",
+  "version": "13.5.1",
   "description": "Architectural sight for AI agents before they edit. Pre-indexes symbols, call graphs, dependencies, architecture layers, and git history into a local SQLite DB. 243 MCP tools, 10 resources, 5 prompts, 28 languages. 100% local, zero API keys.",
   "vendor": {
     "name": "Cranot",

{roam_code-13.5 → roam_code-13.5.1}/src/roam/output/source_context.py

@@ -1,7 +1,7 @@
 """Shared source-context primitives — "code AT a location, with a staleness
 guard."
 
-Why this module exists (2026-06-02 dogfood): the dominant fallback pattern
+Why this module exists (dogfood): the dominant fallback pattern
 in production telemetry is that roam tools return a LOCATION (file:line) and
 the agent then re-greps / re-Reads to see the actual code. We fixed this for
 `roam search` (body_preview) and `roam uses` (call_line) — but with DUPLICATED

{roam_code-13.5 → roam_code-13.5.1}/src/roam/plan/compiler.py

@@ -83,7 +83,7 @@ _STRUCTURAL_DEAD_RE = re.compile(
     re.IGNORECASE,
 )
 _STRUCTURAL_CYCLE_RE = re.compile(
-    # W31 (2026-05-30): Phase A --explain smoke discovered "are there cycles
+    # W31: Phase A --explain smoke discovered "are there cycles
     # in X imports?" fell through to freeform_explore. Pattern was anchored on
     # "cycle.*import" (singular). Now accepts the plural and direct "cycles in".
     r"\b(cyclical|cycles?\s+in\b|cycles?.*import|import.*cycles?|"
@@ -97,7 +97,7 @@ _STRUCTURAL_COMPLEXITY_RE = re.compile(
     # v0.3 fix: dropped "worst-case complexity" — that intent (deep07) needs
     # source reading, not file_info. Let it fall through to freeform_explore
     # which wins on algorithm-source questions.
-    # W31 (2026-05-30): Phase A --explain smoke discovered "top N most complex
+    # W31: Phase A --explain smoke discovered "top N most complex
     # symbols in X" fell through. Added "most complex" / "top N complex".
     r"\b(god[\s-]?components?|cognitive complexity|complexity refactor|"
     r"most complex|top\s+\d+.*complex|too\s+complex|"
@@ -861,7 +861,7 @@ def _extract_symbol_defined_where(task: str) -> str | None:
         # plausible identifiers (`file`, `code`, `that`, `what`, ...).
         if sym.lower() in _W11_STOPWORDS:
             continue
-        # Concept-search guard (2026-06-05 dogfood): "find SQL injection risks",
+        # Concept-search guard (dogfood): "find SQL injection risks",
         # "find security issues", "find performance problems", "find memory
         # leaks" — a PLAIN English noun (not identifier-shaped: no `_`, no digit,
         # all-lower or all-upper) FOLLOWED BY MORE content words is a conceptual
@@ -6994,7 +6994,7 @@ _PROCEDURE_PROBE_SKIPS: dict[str, frozenset[str]] = {
             "entry_points",
         }
     ),
-    # W46 (2026-06-02) — telemetry-driven skips using the ACTUAL labels
+    # W46 — telemetry-driven skips using the ACTUAL labels
     # in _L1_ALWAYS_ON_PROBES (grep_replication / import_audit / compare /
     # pickaxe / criterion / scope_lock / output_shape / conventions /
     # module_name / reachability / config / find_by_desc / why_slow /
@@ -7419,7 +7419,7 @@ def _apply_always_on_extenders(
     return _run_always_on_probes(runnable, task, named_paths, cwd, procedure, prefetched, head)
 
 
-# W42 (2026-06-02): total wall budget across all always_on probes per call.
+# W42: total wall budget across all always_on probes per call.
 # Past this, remaining probes are cancelled. Set generously so we only
 # truncate the long tail. Override via env for tuning.
 _W42_ALWAYS_ON_BUDGET_MS = int(os.environ.get("ROAM_ALWAYS_ON_BUDGET_MS", "2500"))
@@ -8070,7 +8070,7 @@ _ARTIFACT_POLICY = {
     # the probe fires; this policy is the FALLBACK when probe returned no
     # readable frames (file deleted, glob mismatch, etc.).
     "stack_trace_fix": "full",
-    # W11/W12/W13 (2026-06-02): probe data IS the answer (symbol list / top-N
+    # W11/W12/W13: probe data IS the answer (symbol list / top-N
     # ranking / hotspots). "facts" policy minimises envelope size so the agent
     # gets just the probe payload + recommended_first_command, not full
     # structural context. Without these entries, `_ARTIFACT_POLICY.get(p, "full")`
@@ -8129,7 +8129,7 @@ _L1_PROBE_ELIGIBLE = (
     # W181 — refactor_move added; W166 classifier returns it but L1
     # eligibility was missing, silently degrading the envelope.
     "refactor_move",
-    # W11/W12/W13 (2026-06-02) — three new probe families need L1 eligibility
+    # W11/W12/W13 — three new probe families need L1 eligibility
     # so that when their probes fire (and return non-None data), the artifact
     # is labelled `l1_probe` instead of `full`. Without this, the L1 fire-rate
     # KPI under-counted by 46 calls in 2 days. See W22 → compiler-health
@@ -8338,7 +8338,7 @@ def _maybe_append_compile_telemetry(
         "prefetched_keys": keys,
         "envelope_bytes": envelope_bytes,
         "compile_ms": round(compile_ms, 1),
-        # W5 (2026-06-02): stamp agent_mode from env so `compile-stats --by-mode`
+        # W5: stamp agent_mode from env so `compile-stats --by-mode`
         # populates. Host platforms set ROAM_AGENT_MODE in their compile exec env.
         # Rows pre-dating this edit lack the field; `--by-mode` buckets them
         # as 'unknown'.
@@ -8787,7 +8787,7 @@ def compile_for_artifact(plan: "PlanV0", cwd: str | None = None) -> tuple[dict,
     matched-task comparison: -24% turns, -23% cost, +0.6pp HIGHER
     quality, +31% score-per-dollar.
 
-    W33 (2026-05-30): the auto-selector now prefers `to_l1_probe_envelope`
+    W33: the auto-selector now prefers `to_l1_probe_envelope`
     when (a) procedure is structural/trace, (b) named_paths exist, and
     (c) the probe returned procedure-specific facts. This is the "give
     the answer, not the recipe" path — agent receives precomputed
@@ -8834,7 +8834,7 @@ def compile_for_artifact(plan: "PlanV0", cwd: str | None = None) -> tuple[dict,
         or _REFACTOR_MOVE_RE.search(plan.task or "") is not None
         or _W196_LITERAL_RE.search(plan.task or "") is not None  # W196
         or _W201_IMPORT_RE.search(plan.task or "") is not None  # W201
-        # W11/W12/W13 (2026-06-02) — three new probe families are entirely
+        # W11/W12/W13 — three new probe families are entirely
         # probe-driven: the answer IS the probe result (symbol_definitions /
         # top_n_ranking / cli_verb_slow_diagnosis). Promote unconditionally
         # when the policy picked "facts" so the L1 envelope embeds the
@@ -9001,7 +9001,7 @@ def compile_for_artifact(plan: "PlanV0", cwd: str | None = None) -> tuple[dict,
             "structural_complexity": ("complexity_metrics",),
             "structural_cycle": ("cycles", "cycle_count"),
             "trace_query": ("trace_spans",),
-            # W11/W12/W13 (2026-06-02) — procedure-specific probe keys so the
+            # W11/W12/W13 — procedure-specific probe keys so the
             # L1 envelope recognizes that probe data is present and avoids
             # falling back to "full". Without these, the existing
             # `any(k in pre for k in procedure_keys[procedure])` test
@@ -9195,7 +9195,7 @@ def route_for_plan(plan: "PlanV0", cwd: str | None = None, profile_name: str | N
 # Extract explicit file paths mentioned in the task text. Most real tasks
 # name the file(s) they're about — that's a free signal before search-semantic.
 _PATH_RE = re.compile(
-    # W32 (2026-05-30): trailing boundary used to be [$|\s|['"`):.,] which
+    # W32: trailing boundary used to be [$|\s|['"`):.,] which
     # DROPPED paths followed by ? ! ; ] } > etc. — the most common case
     # being natural questions like "what is src/roam/cli.py?" Every prior
     # compile A/B was polluted by this: the obvious path went missing and
@@ -9229,7 +9229,7 @@ _DIR_RE = re.compile(
 def _extract_file_paths(task: str) -> list[str]:
     """Pull file and directory paths from task text. Higher signal than search.
 
-    R10 (2026-05-29): also extracts directory paths like `src/roam/commands/`
+    R10: also extracts directory paths like `src/roam/commands/`
     that are scope anchors even without a specific filename. Empirically
     this cuts ~30% of search-semantic calls (the ones where the user
     referenced a directory but not a specific file inside it).
@@ -9594,7 +9594,7 @@ def _likely_files_from_search(task: str, cwd: str | None, top_n: int = 6) -> tup
     """Hybrid: explicit path mentions first, then symbol-resolution cache,
     then `roam search-semantic` as the fallback.
 
-    R10 (2026-05-29): search-semantic adds tangential noise when the task
+    R10: search-semantic adds tangential noise when the task
     already names files explicitly — skip the subprocess in that case (saves
     ~200ms AND removes the noise).
 
@@ -9716,7 +9716,7 @@ def clear_plan_cache() -> None:
 def compile_plan(task: str, cwd: str | None = None) -> PlanV0:
     """The v0 task compiler. Zero model calls.
 
-    R10 (2026-05-29) speedups:
+    R10 speedups:
     - Skip `roam search-semantic` when explicit paths exist in task text
     - Skip `roam commands` for read-only procedures (structural/trace/freeform);
       only call it when the procedure produces an artifact that needs

{roam_code-13.5 → roam_code-13.5.1}/src/roam/plan/envelope_introspect.py

@@ -1,7 +1,7 @@
 """Canonical compile-envelope introspection — the single source of truth for
 "what is in this envelope / is it a good L1 result."
 
-Why this exists (2026-06-02 dogfood): THREE diagnostic commands
+Why this exists (dogfood): THREE diagnostic commands
 (compiler-corpus, dispatch-trace, and ad-hoc checks) independently re-derived
 "is this L1 / probe-empty / which families." `compiler-corpus` drifted into a
 real bug — it checked ``agent_contract.facts`` when the probe data lives in

{roam_code-13.5 → roam_code-13.5.1}/src/roam/proof_bundle.py

@@ -1,8 +1,7 @@
 """proof_bundle — AgentChangeProofBundle v1 composer.
 
 Reads a legacy `pr-bundle` JSON (at `.roam/pr-bundles/<branch>.json`) and
-produces the AgentChangeProofBundle v1 dict per the schema spec at
-`internal/planning/PROOF-BUNDLE-SCHEMA-2026-05-27.md`.
+produces the AgentChangeProofBundle v1 dict per the schema spec.
 
 Wires together the three already-shipped modules:
   * command_graph (G2 — what CAN be run)

{roam_code-13.5 → roam_code-13.5.1}/src/roam/verdict.py

@@ -1,6 +1,6 @@
 """verdict — closed-enum verdict engine for AgentChangeProofBundle.
 
-Per the proof-bundle schema (`internal/planning/PROOF-BUNDLE-SCHEMA-2026-05-27.md`):
+Per the proof-bundle schema:
 
 Closed verdict enum, every value machine-reason-backed:
 

{roam_code-13.5 → roam_code-13.5.1}/src/roam/verification_contract.py

@@ -22,8 +22,7 @@ Architecture-seal mapping:
   proof_bundle         →  what actually ran + verdict (evidence + verdict — cmd_pr_bundle)
 
 Per the pivot memo (`project_pivot_to_roam_guard`), this is the gap between
-the G2 command_graph and the AgentChangeProofBundle v1 schema. The schema
-spec lives at `internal/planning/PROOF-BUNDLE-SCHEMA-2026-05-27.md`.
+the G2 command_graph and the AgentChangeProofBundle v1 schema.
 """
 
 from __future__ import annotations

{roam_code-13.5 → roam_code-13.5.1/src/roam_code.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: roam-code
-Version: 13.5
+Version: 13.5.1
 Summary: Local codebase intelligence for AI coding agents: structural repo map, change-safety gates, MCP security receipts, and tamper-evident ChangeEvidence packets. 100% local, zero API keys, zero network egress.
 Author: Cranot
 License-Expression: Apache-2.0

{roam_code-13.5 → roam_code-13.5.1}/tests/test_bench_ground_truth.py

@@ -15,6 +15,13 @@ from click.testing import CliRunner
 
 from roam.cli import cli
 from roam.commands import cmd_bench
+from tests._helpers.repo_root import repo_root
+
+# The benchmark oracles live under the gitignored `internal/` and are absent in
+# a clean public checkout / CI; skip oracle-importing tests when they aren't
+# present so the public suite stays green without them.
+_HAS_INTERNAL = (repo_root() / "internal" / "benchmarks" / "oracle_fix_bug.py").exists()
+_requires_internal = pytest.mark.skipif(not _HAS_INTERNAL, reason="internal/ benchmark oracles absent")
 
 # ----- artifact extraction helpers -----
 
@@ -68,6 +75,7 @@ def test_classify_task_shape_other() -> None:
 # ----- oracle dispatch -----
 
 
+@_requires_internal
 def test_ground_truth_score_pytest_dispatch(monkeypatch: pytest.MonkeyPatch) -> None:
     """When shape=write_pytest and a test block exists, oracle_pytest is called."""
     calls: list[dict] = []
@@ -88,6 +96,7 @@ def test_ground_truth_score_pytest_dispatch(monkeypatch: pytest.MonkeyPatch) ->
     assert "def test_foo" in calls[0]["source"]
 
 
+@_requires_internal
 def test_ground_truth_score_pytest_fail(monkeypatch: pytest.MonkeyPatch) -> None:
     def fake_run_produced_test(source, project_root, *, timeout=30):  # noqa: ANN001
         return {"exit_code": 1, "stdout_tail": "", "stderr_tail": "", "duration_ms": 1, "timed_out": False}
@@ -102,6 +111,7 @@ def test_ground_truth_score_pytest_fail(monkeypatch: pytest.MonkeyPatch) -> None
     assert score == "0"
 
 
+@_requires_internal
 def test_ground_truth_score_fix_bug_dispatch(monkeypatch: pytest.MonkeyPatch) -> None:
     """When shape=stack_trace_fix and a diff exists, oracle_fix_bug is called."""
     calls: list[dict] = []
@@ -161,6 +171,7 @@ def test_bench_compile_help_shows_ground_truth_flag() -> None:
     assert "--ground-truth" in res.output
 
 
+@_requires_internal
 def test_bench_compile_default_off_no_oracle_calls(
     monkeypatch: pytest.MonkeyPatch,
     tmp_path: Path,
@@ -223,6 +234,7 @@ def test_bench_compile_default_off_no_oracle_calls(
     assert bug_calls == []
 
 
+@_requires_internal
 def test_bench_compile_ground_truth_records_score(
     monkeypatch: pytest.MonkeyPatch,
     tmp_path: Path,
@@ -339,16 +351,8 @@ def test_bench_compile_ground_truth_unsupported_shape_empty_score(
 
 # ----- cwd-decoupling regression (SWE-bench readiness) -----
 
-# The benchmark oracles live under the gitignored `internal/` and are absent in
-# a clean public checkout / CI; skip these when they aren't present so the
-# public suite stays green without them. Use the repo_root helper (W588) rather
-# than a fragile Path(__file__) walk.
-from tests._helpers.repo_root import repo_root  # noqa: E402
-
-_HAS_INTERNAL = (repo_root() / "internal" / "benchmarks" / "oracle_fix_bug.py").exists()
-
 
-@pytest.mark.skipif(not _HAS_INTERNAL, reason="internal/ benchmark oracles absent")
+@_requires_internal
 def test_ensure_internal_importable_true() -> None:
     """The dev-repo root is derivable from roam.__file__ so internal/ resolves."""
     assert cmd_bench._ensure_internal_importable() is True

{roam_code-13.5 → roam_code-13.5.1}/tests/test_cmd_compile.py

@@ -1,8 +1,8 @@
 """Tests for `roam compile` — freeform task → structured envelope.
 
 Empirically validated 2026-05-28 (Opus 4.8 spike): facts envelope
-delivers 99% of vanilla quality at 54% of cost. See
-internal/planning/FINAL-EVALUATION-2026-05-28.md.
+delivers 99% of vanilla quality at 54% of cost.
+
 """
 
 from __future__ import annotations
@@ -44,7 +44,7 @@ def test_compile_json_mode_basic(runner):
 def test_compile_structural_routes_to_full(runner):
     """structural_* procedures route to `full` OR `l1_probe`.
 
-    W33 (2026-05-30): structural tasks with named paths now prefer the
+    W33: structural tasks with named paths now prefer the
     l1_probe envelope when the probe returns procedure-specific facts
     (the agent gets the answer, not just metadata). Either is the right
     choice; the test accepts both since the index may not always have

{roam_code-13.5 → roam_code-13.5.1}/tests/test_compile_describe_module.py

@@ -1,4 +1,4 @@
-"""W1 (2026-06-09) — module-level describe recall.
+"""W1 — module-level describe recall.
 
 Telemetry: "explain the compiler architecture" / "what does the constitution
 package do" / "what are the public functions exported by X" leaked to

{roam_code-13.5 → roam_code-13.5.1}/tests/test_compile_probe_families_w11_w12_w13.py

@@ -29,7 +29,7 @@ from roam.plan.compiler import (
 
 _W11_POSITIVE = [
     "where is _evaluate_mcp_mode_policy defined",
-    "find where useKiniseisBalance is",
+    "find where useLedgerBalance is",
     "where is compile_plan defined",
     "which file defines RoamPluginContext",
     "where is run_produced_test",
@@ -54,7 +54,7 @@ _W11_NEGATIVE = [
     "where is the bug in src/roam/cli.py",
     # generic English question — no concrete bareword
     "where is the documentation",
-    # W11.5 (2026-06-05 dogfood) — concept searches: a PLAIN noun followed by
+    # W11.5 (dogfood) — concept searches: a PLAIN noun followed by
     # more words is conceptual, NOT "find symbol X" (was extracting SQL/security
     # /performance/memory as bogus symbols → garbage prefetch).
     "find SQL injection risks",
@@ -315,7 +315,7 @@ def test_extractor_helpers_consistent_with_classifier() -> None:
         assert _is_cli_verb_why_slow(task), f"W13 helper missed {task!r}"
 
 
-# ---- cli-command file resolution (2026-06-05 dogfood) ----
+# ---- cli-command file resolution (dogfood) ----
 
 
 def test_resolve_cli_command_files():

{roam_code-13.5 → roam_code-13.5.1}/tests/test_compile_stats_by_mode.py

@@ -1,4 +1,4 @@
-"""W5 (2026-06-02) — tests for the new ``--by-mode`` flag on
+"""W5 — tests for the new ``--by-mode`` flag on
 ``roam compile-stats``.
 
 The flag groups telemetry rows by ``agent_mode`` (compile/roam/vanilla/

{roam_code-13.5 → roam_code-13.5.1}/tests/test_mcp_server_card_hash.py

@@ -39,7 +39,7 @@ from tests._helpers.repo_root import repo_root
 # og.png pointing at deployed assets on roam-code.com). All 3
 # .well-known card path variants stay byte-identical per the W792
 # invariant.
-_EXPECTED_CARD_SHA256 = "4b81ef9dca72a84cbde8776838adb4424d0c7782fca0191f1fc0f8e6baf2eca4"
+_EXPECTED_CARD_SHA256 = "d78404f85270c32a949c2e9f049531d61701955247f2029cd597f3fe3affb6ae"
 
 
 def _card_path() -> Path:

{roam_code-13.5 → roam_code-13.5.1}/tests/test_plan_artifact_selector.py

@@ -185,7 +185,7 @@ def test_select_artifact_structural_routes_to_full():
 def test_compile_for_artifact_returns_correct_schema():
     """compile_for_artifact returns (envelope, label) tuple.
 
-    W33 (2026-05-30): when probe fires for a structural task with a named
+    W33: when probe fires for a structural task with a named
     path, the selector now prefers `l1_probe` over `full` (it contains the
     precomputed answer, not just metadata). Either label is acceptable; the
     schema reflects which one was picked.