attocode-code-intel 0.1.13__tar.gz → 0.1.15__tar.gz

{attocode_code_intel-0.1.13 → attocode_code_intel-0.1.15}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: attocode-code-intel
-Version: 0.1.13
+Version: 0.1.15
 Summary: Local code search and navigation for coding agents and people
 Project-URL: Homepage, https://github.com/eren23/attocode
 Project-URL: Repository, https://github.com/eren23/attocode
@@ -130,6 +130,8 @@ attocode-code-intel init
 Then restart your agent. `init` configures Codex, Claude Code and Cursor. Use `--client claude`, `--client codex` or `--client cursor` for one client. Codex needs a trusted project for a project configuration. Claude Code can ask you to enable the project MCP server.
 
 To check the setup, run `attocode-code-intel doctor`. It starts the configured server, calls two tools and shows the result for each client.
+Development builds also report the executable and server version returned by
+that new probe. An already running client session must be checked separately.
 
 For other clients (VS Code, Zed, Windsurf, Claude Desktop and more), run `attocode-code-intel install <target>`. `attocode-code-intel commands` lists the targets.
 
@@ -149,7 +151,11 @@ tests, such as `_test.dart` and Perl `.t` scripts and `t/` modules. Version
 0.1.12 makes `doctor --client claude` read the local-scope entry that
 `claude mcp add` writes to `~/.claude.json`. Version 0.1.13 fills compact search
 pages within their budget, offers bounded deeper searches, and repairs stale
-navigation indexes after successful restart parsing. See the [change investigation guide](https://github.com/eren23/attocode/blob/main/docs/guides/change-investigation.md).
+navigation indexes after successful restart parsing. Version 0.1.15 adds parser ownership by thread, one language-server stderr
+reader, bounded native recovery without static callers and release install
+verification on GitHub runners. Version 0.1.14 adds bounded
+selected-file recovery, safe navigation publication, guarded Elixir definitions
+and doctor probe identity. See the [change investigation guide](https://github.com/eren23/attocode/blob/main/docs/guides/change-investigation.md).
 
 The agent sees 16 tools (the `daily` profile):
 
@@ -199,6 +205,9 @@ Search matches the words of the code and its paths. It does not use embeddings u
 The [R6 review study](https://github.com/eren23/attocode/blob/main/packages/code-intel/evals/results/2026-10-13-r6-review-test-note.md)
 found no demonstrated review benefit. The [completed R7 repair pilot](https://github.com/eren23/attocode/blob/main/packages/code-intel/evals/results/2026-10-14-r7-completion-pilot.md)
 also found no demonstrated benefit; its reviewed decision is to stop expansion.
+The [2026-10-22 investigation pilot](https://github.com/eren23/attocode/blob/main/packages/code-intel/evals/results/2026-10-22-investigation-pilot.md)
+stopped incomplete after its first task pair because structured-output/tool calls
+failed the frozen protocol. It provides no quality, cost or time benefit evidence.
 
 ## Limits and troubleshooting
 
@@ -206,6 +215,11 @@ also found no demonstrated benefit; its reviewed decision is to stop expansion.
 - **Index location.** The index is in `.attocode/` in your project. When the server makes that folder, it adds a `.gitignore`, so the index stays out of commits.
 - **What the index reads.** The server reads at most 50,000 files and skips files over 1 MB. It skips files that `.gitignore` ignores, and folders that start with a dot. It also skips folders named `dist`, `build`, `out`, `output`, `target`, `env`, `tmp`, `vendor` or `coverage`, and binary, media and lock files. If a build-named folder contains real source, explicitly re-include it with a root `.gitignore` entry such as `!dist/`; discovery and edit refresh honor that override.
 - **Languages.** The base install parses Bash, C, C#, C++, CSS, Elixir, Go, HCL, HTML, Java, JavaScript, JSON, Kotlin, Lua, PHP, Python, Ruby, Rust, Scala, Swift, TOML, TypeScript, YAML and Zig. Optional grammars are loaded only when already installed or cached; installing the `tree-sitter-all` extra does not trigger grammar downloads during navigation. A slow grammar falls back after a 100 ms parse budget. Static imports and references are partial evidence, not compiler/type resolution.
+- **Degraded parsing (development builds).** Selecting a missing definition in
+  a file with a recorded native parse failure permits one bounded retry for
+  that source/parser generation. A failed retry reports partial evidence and
+  an `attocode-code-intel notify` command to queue a file refresh. Empty results
+  and ordinary syntax errors do not trigger this recovery.
 - **Test paths.** Search and test suggestions recognize conventional test folders and filenames, plus supported Kotlin/Android source sets directly under `src` (for example `commonTest`, `jvmTest`, `androidHostTest` and `androidDeviceTest`). Helpers in those sets count as test code even without a test filename. Custom target or flavor names are not inferred from a `Test` suffix alone. Short queries explicitly asking for tests put recognized test files first, preserving order within each group; longer issue descriptions keep source-first behavior. Suggestions can include helpers; verify the actual runnable tests.
 - **Remove the setup.** Run `attocode-code-intel init --remove`. It restores an entry that existed before `init`.
 - **Server does not answer.** Run `attocode-code-intel doctor`. If a client shows "not configured", run `init` for it. After a change, restart the agent. In Claude Code, `/mcp` shows the server state.

{attocode_code_intel-0.1.13 → attocode_code_intel-0.1.15}/README.md

@@ -17,6 +17,8 @@ attocode-code-intel init
 Then restart your agent. `init` configures Codex, Claude Code and Cursor. Use `--client claude`, `--client codex` or `--client cursor` for one client. Codex needs a trusted project for a project configuration. Claude Code can ask you to enable the project MCP server.
 
 To check the setup, run `attocode-code-intel doctor`. It starts the configured server, calls two tools and shows the result for each client.
+Development builds also report the executable and server version returned by
+that new probe. An already running client session must be checked separately.
 
 For other clients (VS Code, Zed, Windsurf, Claude Desktop and more), run `attocode-code-intel install <target>`. `attocode-code-intel commands` lists the targets.
 
@@ -36,7 +38,11 @@ tests, such as `_test.dart` and Perl `.t` scripts and `t/` modules. Version
 0.1.12 makes `doctor --client claude` read the local-scope entry that
 `claude mcp add` writes to `~/.claude.json`. Version 0.1.13 fills compact search
 pages within their budget, offers bounded deeper searches, and repairs stale
-navigation indexes after successful restart parsing. See the [change investigation guide](https://github.com/eren23/attocode/blob/main/docs/guides/change-investigation.md).
+navigation indexes after successful restart parsing. Version 0.1.15 adds parser ownership by thread, one language-server stderr
+reader, bounded native recovery without static callers and release install
+verification on GitHub runners. Version 0.1.14 adds bounded
+selected-file recovery, safe navigation publication, guarded Elixir definitions
+and doctor probe identity. See the [change investigation guide](https://github.com/eren23/attocode/blob/main/docs/guides/change-investigation.md).
 
 The agent sees 16 tools (the `daily` profile):
 
@@ -86,6 +92,9 @@ Search matches the words of the code and its paths. It does not use embeddings u
 The [R6 review study](https://github.com/eren23/attocode/blob/main/packages/code-intel/evals/results/2026-10-13-r6-review-test-note.md)
 found no demonstrated review benefit. The [completed R7 repair pilot](https://github.com/eren23/attocode/blob/main/packages/code-intel/evals/results/2026-10-14-r7-completion-pilot.md)
 also found no demonstrated benefit; its reviewed decision is to stop expansion.
+The [2026-10-22 investigation pilot](https://github.com/eren23/attocode/blob/main/packages/code-intel/evals/results/2026-10-22-investigation-pilot.md)
+stopped incomplete after its first task pair because structured-output/tool calls
+failed the frozen protocol. It provides no quality, cost or time benefit evidence.
 
 ## Limits and troubleshooting
 
@@ -93,6 +102,11 @@ also found no demonstrated benefit; its reviewed decision is to stop expansion.
 - **Index location.** The index is in `.attocode/` in your project. When the server makes that folder, it adds a `.gitignore`, so the index stays out of commits.
 - **What the index reads.** The server reads at most 50,000 files and skips files over 1 MB. It skips files that `.gitignore` ignores, and folders that start with a dot. It also skips folders named `dist`, `build`, `out`, `output`, `target`, `env`, `tmp`, `vendor` or `coverage`, and binary, media and lock files. If a build-named folder contains real source, explicitly re-include it with a root `.gitignore` entry such as `!dist/`; discovery and edit refresh honor that override.
 - **Languages.** The base install parses Bash, C, C#, C++, CSS, Elixir, Go, HCL, HTML, Java, JavaScript, JSON, Kotlin, Lua, PHP, Python, Ruby, Rust, Scala, Swift, TOML, TypeScript, YAML and Zig. Optional grammars are loaded only when already installed or cached; installing the `tree-sitter-all` extra does not trigger grammar downloads during navigation. A slow grammar falls back after a 100 ms parse budget. Static imports and references are partial evidence, not compiler/type resolution.
+- **Degraded parsing (development builds).** Selecting a missing definition in
+  a file with a recorded native parse failure permits one bounded retry for
+  that source/parser generation. A failed retry reports partial evidence and
+  an `attocode-code-intel notify` command to queue a file refresh. Empty results
+  and ordinary syntax errors do not trigger this recovery.
 - **Test paths.** Search and test suggestions recognize conventional test folders and filenames, plus supported Kotlin/Android source sets directly under `src` (for example `commonTest`, `jvmTest`, `androidHostTest` and `androidDeviceTest`). Helpers in those sets count as test code even without a test filename. Custom target or flavor names are not inferred from a `Test` suffix alone. Short queries explicitly asking for tests put recognized test files first, preserving order within each group; longer issue descriptions keep source-first behavior. Suggestions can include helpers; verify the actual runnable tests.
 - **Remove the setup.** Run `attocode-code-intel init --remove`. It restores an entry that existed before `init`.
 - **Server does not answer.** Run `attocode-code-intel doctor`. If a client shows "not configured", run `init` for it. After a change, restart the agent. In Claude Code, `/mcp` shows the server state.

attocode_code_intel-0.1.15/evals/results/2026-10-21-test-case-spans-decision.json

@@ -0,0 +1,140 @@
+{
+  "status": "rejected",
+  "checks": {
+    "two_new_first_page_matches_every_round": false,
+    "gains_in_both_repositories_every_round": false,
+    "no_file_or_excerpt_loss": false,
+    "controls_identical": true,
+    "warm_p95_each_round": true,
+    "cleanup_verified": true
+  },
+  "gains": [],
+  "losses": [
+    {
+      "id": "express:fresh-test-spans-01",
+      "round": 0,
+      "depth": 5,
+      "kind": "files",
+      "label": "test/res.status.js"
+    },
+    {
+      "id": "express:fresh-test-spans-01",
+      "round": 1,
+      "depth": 5,
+      "kind": "files",
+      "label": "test/res.status.js"
+    },
+    {
+      "id": "express:fresh-test-spans-01",
+      "round": 2,
+      "depth": 5,
+      "kind": "files",
+      "label": "test/res.status.js"
+    },
+    {
+      "id": "express:fresh-test-spans-04",
+      "round": 0,
+      "depth": 5,
+      "kind": "files",
+      "label": "test/req.get.js"
+    },
+    {
+      "id": "express:fresh-test-spans-04",
+      "round": 1,
+      "depth": 5,
+      "kind": "files",
+      "label": "test/req.get.js"
+    },
+    {
+      "id": "express:fresh-test-spans-04",
+      "round": 2,
+      "depth": 5,
+      "kind": "files",
+      "label": "test/req.get.js"
+    },
+    {
+      "id": "express:fresh-test-spans-11",
+      "round": 0,
+      "depth": 5,
+      "kind": "files",
+      "label": "test/res.sendStatus.js"
+    },
+    {
+      "id": "express:fresh-test-spans-11",
+      "round": 1,
+      "depth": 5,
+      "kind": "files",
+      "label": "test/res.sendStatus.js"
+    },
+    {
+      "id": "express:fresh-test-spans-11",
+      "round": 2,
+      "depth": 5,
+      "kind": "files",
+      "label": "test/res.sendStatus.js"
+    },
+    {
+      "id": "express:fresh-test-spans-12",
+      "round": 0,
+      "depth": 5,
+      "kind": "files",
+      "label": "test/app.head.js"
+    },
+    {
+      "id": "express:fresh-test-spans-12",
+      "round": 1,
+      "depth": 5,
+      "kind": "files",
+      "label": "test/app.head.js"
+    },
+    {
+      "id": "express:fresh-test-spans-12",
+      "round": 2,
+      "depth": 5,
+      "kind": "files",
+      "label": "test/app.head.js"
+    }
+  ],
+  "p95_seconds": {
+    "0": {
+      "baseline": 0.0862042079679668,
+      "candidate": 0.07823979202657938
+    },
+    "1": {
+      "baseline": 0.08303270791657269,
+      "candidate": 0.08827808301430196
+    },
+    "2": {
+      "baseline": 0.07845087500754744,
+      "candidate": 0.07912220794241875
+    }
+  },
+  "counts": {
+    "5": {
+      "baseline": {
+        "files": 15,
+        "excerpts": 4
+      },
+      "candidate": {
+        "files": 11,
+        "excerpts": 6
+      }
+    },
+    "24": {
+      "baseline": {
+        "files": 15,
+        "excerpts": 4
+      },
+      "candidate": {
+        "files": 16,
+        "excerpts": 11
+      }
+    }
+  },
+  "evidence_scope": "consumed_development",
+  "fresh_validation_performed": false,
+  "initial_attempt": "operator_invalid; zero callback units; not relevance evidence",
+  "candidate_rule_unchanged": true,
+  "archive_sha256": "c2bc98099fe2e66b7815c75ecf5395ba436ac73ade2088507d79630987df79b4",
+  "questions_sha256": "c09fd7845c4cd0477b64d60debd4ff7c523f837123f71f992557022432c3b370"
+}

attocode_code_intel-0.1.15/evals/results/2026-10-21-test-case-spans.md

@@ -0,0 +1,78 @@
+# JavaScript/TypeScript test-case spans — rejected on development
+
+The initial comparison is **operator-invalid**, not a test of the intended
+callback-span hypothesis. macOS workspace aliases (`/var` versus `/private/var`)
+caused the adapter to fail before extraction. The index caught that exception
+and used file windows. All 2,088 candidate hit occurrences were file windows;
+there were zero callback units. The earlier arithmetic rejection must not be
+interpreted as evidence that correctly indexed callback spans are ineffective.
+
+An independent author and reviewer grounded 24 new questions in pinned Express
+and Prisma sources before that attempt. The questions are now consumed. Any
+corrected comparison on them is development evidence, not fresh validation.
+The initial failed source-copy attempt, fallback-only comparison, raw pages,
+operators and earlier uncommitted report/archive are retained privately.
+
+The technical correction canonicalizes workspace and AST paths. An alias-path
+regression fails before the fix and passes afterward. Before any query, the
+operator requires actual indexed callback windows, identical keyword-document
+hashes, and identical full function/method-window hashes. Class fallback window
+counts can legitimately change when a callback owner replaces that source span;
+an earlier overly strict admission check stopped Prisma before querying and is
+also retained. The extraction rule, scoring configuration and questions did not
+change. **No search ranking change is enabled in product.**
+
+## Corrected, admitted development comparison
+
+The rule recognizes literal `it`/`test` callbacks and single `.only`/`.skip`
+forms in JavaScript/TypeScript test paths. Dynamic titles, arbitrary receivers
+and parameterized DSLs are not inferred. It adds search-only body units, without
+creating navigation definitions. Only short explicit test queries use the
+alternate index; ordinary queries use the original index.
+
+The corrected comparison ran three alternating rounds through actual compact
+MCP delivery, requesting depths 5 and 24 with 2,000 and 8,000 token budgets.
+Admission found 1,127 test-case windows from 1,119 distinct owners in Express,
+and 4,809 windows from 3,935 owners in Prisma. Keyword and function/method-window
+hashes match their baseline in each repository. The delivered results include
+1,167 test-case hit occurrences, unlike the invalid initial attempt.
+
+| Delivered evidence | Baseline | Candidate |
+|---|---:|---:|
+| Labeled query/file matches at 5 | 15 | 11 |
+| Labeled query/file matches at 24 | 15 | 16 |
+| Labeled excerpts at 5 | 4 | 6 |
+| Labeled excerpts at 24 | 4 | 11 |
+
+Counts are identical in all three rounds. There are zero new first-page file
+matches. Express status validation, referrer lookup, `sendStatus`, and HEAD
+fallback each lose a labeled file at five in every round: 12 loss observations
+over four questions. Extra excerpt matches do not satisfy the frozen no-loss
+rule. **Reject on consumed development; skip new fresh validation and do not
+retune this rule.** This is not a held-out relevance gain or general search-quality
+claim.
+
+Warm p95 was 86.2/83.0/78.5 ms for baseline and 78.2/88.3/79.1 ms for candidate.
+All three latency checks pass. All six ordinary-query raw control pairs match.
+The two admitted workers took 27.609 seconds, with peak sampled RSS 188.3 MiB
+and peak managed storage 135.3 MiB. Child cleanup and owned scratch removal pass.
+No model, source download or cloud worker was used for this comparison.
+
+## Replay and limits
+
+The [decision](2026-10-21-test-case-spans-decision.json) binds the
+[archive](2026-10-21-test-case-spans.tar.gz). Extract it outside the checkout and
+run its `replay.py` with this repository and `packages/code-intel/src` on
+`PYTHONPATH`. It verifies file hashes, recomputes all 288 admitted page judgments,
+checks controls and representation admissions, and reproduces every gate without
+retrieval or network access. The earlier invalid attempt is explicitly segregated
+inside the archive and does not contribute to the development decision.
+
+Public envelopes retain result data; private workspace metadata is omitted and
+control follow-up roots use `$SOURCE_ROOT`. Source snippets, coordinates, ordering,
+timings and judgments are unchanged; original response hashes are recorded.
+Independent review checked 39 pinned label files, all request/order bindings,
+835 coordinate-bearing delivered windows including controls, and all admitted
+test-case hit identities. One metadata preview has no source coordinates and
+never earns excerpt credit. Labels are source-bound and non-exhaustive; no agent
+completion benefit is established.

attocode_code_intel-0.1.15/evals/results/2026-10-22-empty-caller-recovery.json

@@ -0,0 +1,84 @@
+{
+  "base": "9c09cc8e4aa1b38812f1bb09413b7d42b8e66d61",
+  "ci_requirements": {
+    "precision_recovery": 11,
+    "real_server_workflows": 4,
+    "relevant_regressions": 165,
+    "sql_per_python": 51
+  },
+  "date": "2026-10-22",
+  "evidence_sha256": {
+    "integration/actual-inventory-audit.json": "f54d1018785da00fd0f4799878ec93bdea054737b32bb2406f19a0e41bf2e900",
+    "integration/precision-gate-proof.json": "fdfabf210ed2bab28c2e9c7843b2374366bd59157a578efb569e3de42ea8cb3c",
+    "local/contracts.xml": "27a028ba5d119719934f3b3fae5ea9ee486b1cd29f618a95da818269e9b3f946",
+    "local/runtime.json": "163359e0038d1ee2eb5bc643aa2dc92a8c6c84376b38284bb83a4a004b5958b0",
+    "ocaml/summary.json": "612373cd31b624c5bf76026642f2d40c9fc1f5dcd777ccd02b2b82d7ff937c15",
+    "review/final-baseline.xml": "7b5094ee100b3624d7b60cd13483b269db9c4eb84b8af4428b5f79137ae7b8f3",
+    "review/final-green.xml": "9cbbaed25ed0904262fa1e4a5ed0eead95ded2dbbbafb9679e585cfb0ea7df95",
+    "review/review-final.json": "54fef6da97f6dd102243066f8deb737fe496f207cb98aa24dff41b46a0b5f3e7"
+  },
+  "existing_local_verification": {
+    "expected_failures": 1,
+    "owned_scratch_removed": true,
+    "passed": 4157,
+    "peak_managed_disk_bytes": 563075651,
+    "peak_rss_bytes": 981565440,
+    "seconds": 162.7486903751269,
+    "skipped": 14,
+    "sql_passed": 50,
+    "sql_skipped": 1
+  },
+  "limits": [
+    "Historical Rust and OCaml triggers remain unproven",
+    "No ranking or Cockroach restart-speed gain",
+    "Remaining validation runs in GitHub CI; no further local test suite",
+    "No release, installation change or host MCP restart"
+  ],
+  "ocaml": {
+    "conclusion": "Both current observations pass; no native timeout, extraction loss or definition publication loss was observed. Wall and threadCPU are close for these three successful native calls. No new OCaml product patch is justified by these observations.",
+    "historical_limit": "Archived natural-1 already loses the native tree before extraction, consistent with exhaustion of the100ms native budget. Its exception/threadCPU/scheduling data are unavailable. New successful observations cannot establish the historical trigger or attribute it to scheduling, page-in, extraction, cache publication or the earlier parser-thread defect.",
+    "native_calls": [
+      {
+        "caller": "ts_parse_file",
+        "effective_limit_micros": 100000,
+        "mode": "direct",
+        "native_outcome": "tree",
+        "source_sha256": "c3abd2af1040c9d1ad305f76b7f6f9f2e8243d45af2b862b10906ee7bd629010",
+        "thread_cpu_micros": 20063,
+        "wall_micros": 20077
+      },
+      {
+        "caller": "ts_parse_file",
+        "effective_limit_micros": 100000,
+        "mode": "gateway",
+        "native_outcome": "tree",
+        "source_sha256": "c3abd2af1040c9d1ad305f76b7f6f9f2e8243d45af2b862b10906ee7bd629010",
+        "thread_cpu_micros": 18521,
+        "wall_micros": 18526
+      },
+      {
+        "caller": "syntax_tree",
+        "effective_limit_micros": 100000,
+        "mode": "gateway",
+        "native_outcome": "tree",
+        "source_sha256": "c3abd2af1040c9d1ad305f76b7f6f9f2e8243d45af2b862b10906ee7bd629010",
+        "thread_cpu_micros": 18605,
+        "wall_micros": 18626
+      }
+    ],
+    "seconds": 4.871503083966672,
+    "workers": 2
+  },
+  "regression": {
+    "attempt_limit": 4,
+    "baseline_failed": 4,
+    "baseline_ready_unused_passed": 1,
+    "deadline_seconds": 2,
+    "fixed_passed": 5
+  },
+  "scope": "Continue existing bounded recovery when both native lookups are unresolved and static callers are absent",
+  "sources": {
+    "packages/code-intel/src/attocode_intel/precision.py": "7245a019ab647448839c12bbd3043ce3d96f38640512c9ef61c75d32070cd165",
+    "packages/code-intel/tests/test_precision_recovery.py": "cc544b561735510fe168ffbfa161434070143d177b0d35ede9ccaebd7ac5be7e"
+  }
+}

attocode_code_intel-0.1.15/evals/results/2026-10-22-empty-caller-recovery.md

@@ -0,0 +1,46 @@
+# Precision recovery without static callers
+
+Precision enrichment previously stopped after unresolved definition and reference
+responses when static analysis found no callers. Those responses can be transient
+even after a successful server warmup. Recovery now continues within the existing
+two-second deadline and four-attempt cap. A resolved unused declaration still
+takes the fast path; persistent null responses remain partial.
+
+The [machine report](2026-10-22-empty-caller-recovery.json) records source and
+evidence hashes. Five gateway regressions use a framed subprocess peer after a
+successful external definition warmup. They verify recovery with and without a
+dynamic caller, a resolved unused declaration, persistent nulls, deadline expiry,
+cached outcomes and shutdown. Static callers are actually absent. Against the
+base implementation, four behavior cases fail and the ready-unused control passes;
+all five pass after the fix. The cache clock is frozen independently of the real
+asyncio deadline to avoid scheduler-dependent cooldown assertions.
+
+One natural Rust observation before the fix passed in 5.480 seconds and had one
+static caller, so it did not exercise this edge. A later installed Rust workflow
+also passed. Neither observation explains the historical intermittent Rust failure.
+An earlier observer setup import error started no server; an initial regression
+fixture accidentally supplied a static import reference. Both unsuccessful setup
+attempts remain in the private evidence and are excluded from behavioral proof.
+
+The existing full local result is 4,157 passed, 14 skipped and one expected failure.
+Its source-bound controller completed in 162.749 seconds, with sampled peak RSS
+936.1 MiB and managed scratch 537.0 MiB; owned scratch was removed. Lint, scale at
+100/1,000/10,000 files and 20 lifecycle cycles passed. Local SQL checks were 50
+passed and one PostgreSQL-dependent skip; CI must independently pass all 51 on
+both supported Python versions. These local runs preceded the user's correction
+to keep broad testing online. No additional local test run is authorized by this
+report; remaining validation uses GitHub CI.
+
+The separate OCaml check used two predeclared fresh processes on the exact
+historical source. Three native calls returned clean trees in 18.5–20.1 ms with
+similar thread CPU times under the unchanged 100 ms limit. The initial AST and
+published index already contained the target before selected inspection. No
+OCaml patch follows from these successful observations. Historical no-tree
+results occurred before extraction, but their suppressed exception and absent
+CPU measurements prevent attribution to a specific trigger. Direct execution
+preceded the gateway worker, so filesystem and library caches may have warmed.
+
+Independent review found no remaining implementation issue. The CI evidence gate
+additionally requires all 11 precision-recovery identities alongside the prior
+154 regressions, all SQL cases and all four real-server workflows. No publication,
+installation change, MCP restart or ranking improvement is claimed.

attocode_code_intel-0.1.15/evals/results/2026-10-22-investigation-pilot-decision.json

@@ -0,0 +1,61 @@
+{
+  "archive": {
+    "file": "2026-10-22-investigation-pilot.tar.gz",
+    "sha256": "3ddbd267583d793ef3153e88760ae5aaebe751501e2cde7f1d7ba9530324dd24"
+  },
+  "billing": "Subscription extra usage was off at preflight; reported API-equivalent cost is not an invoice. Interrupted native cost unknown.",
+  "cleanup_verified": true,
+  "cost_ratio": null,
+  "decision": "stop_without_retry_or_replacement",
+  "excluded_wiring_attempts": 2,
+  "exposure": {
+    "novelty_limit": "I07 composite primary was new; one supporting caller/function anchor was previously consumed and disclosed before launch.",
+    "other_frozen_tasks_launched": false,
+    "scored_tasks_exposed_to_client": [
+      "I07"
+    ],
+    "unlaunched_prompts_and_labels_exported": false
+  },
+  "fully_correct_delta": null,
+  "known_reported_cost_subtotal_usd": "0.1823414",
+  "no_agent_benefit_claim": true,
+  "observed_scored_attempts": 2,
+  "operators": {
+    "guard.py": "cbb68e2abafb776b5561007bb79f8fdfaf52f0902cb152031553e80d76192692",
+    "pilot.py": "2be1331af2a0276382591ea9663904c86e61592784d098ca5f4f4df4e03f4d58",
+    "proxy.py": "1ed9397130d4cb04d56884245887050945a260ad4d7aeea85344d37020f1c8ad"
+  },
+  "planned_scored_attempts": 16,
+  "release_source_head": "08319af02eeee95b88978dd126e392fec67719de",
+  "replay_scope": "Sanitized accounting and structural tool-call evidence only. No raw answer, semantic grade or causal agent-benefit estimate.",
+  "repositories": {
+    "cli/cli": {
+      "revision": "6e49747f49f31c9c0a68cfa344df7262a684e723",
+      "sha256": "fcd228aeb35f49ec460788314c482586f1da41a4a6ebaf3632261ace87fd7468",
+      "snapshot": "sources/repo-0"
+    },
+    "fastapi/fastapi": {
+      "revision": "11614be9021aa4ac078d4d0693a8b5250a1010d8",
+      "sha256": "17063c3d3d1b1b26257e17b16128b3d1fa6f4d90cb7e1a6c93b7b62604b5fbe6",
+      "snapshot": "sources/repo-1"
+    },
+    "phoenixframework/phoenix": {
+      "revision": "850162d618cf4e31be82fb53ad6e259095fa6126",
+      "sha256": "200f91bae027b3f9d078c9a2e32795440b8fd778b5ac96b21726ba03177294db",
+      "snapshot": "sources/repo-2"
+    },
+    "psf/requests": {
+      "revision": "ebf71906798ec82f34e07d3168f8b8aecaf8a3be",
+      "sha256": "63efe90377f671ce0bfa71737de62e7545b6ed4177458033251843776023d0c4",
+      "snapshot": "sources/repo-3"
+    }
+  },
+  "schema": 1,
+  "semantic_grades": null,
+  "status": "incomplete",
+  "stop_reason": "Malformed structured-output/tool calls; frozen protocol failure cancels the active pair.",
+  "study_freeze_sha256": "b6bd8224dea8ff91e2bca7fbcd5690162290002341ad339eaa6d9b5006d47d23",
+  "time_ratio": null,
+  "total_reported_cost_usd": null,
+  "unlaunched_scored_attempts": 14
+}

attocode_code_intel-0.1.15/evals/results/2026-10-22-investigation-pilot.md

@@ -0,0 +1,108 @@
+# Investigation pilot: incomplete, expansion stopped
+
+The conditional investigation pilot stopped after its first task pair because
+structured-output/tool calls failed the frozen protocol. Two scored attempts
+were admitted; fourteen were never launched. Neither scored attempt completed
+the study gates. No quality, cost or time benefit is established, and beta
+positioning remains unchanged. The earlier R7 stop decision remains closed.
+
+The [machine decision](2026-10-22-investigation-pilot-decision.json) and
+[sanitized evidence archive](2026-10-22-investigation-pilot.tar.gz) retain the
+incomplete outcomes. Archive SHA-256:
+`3ddbd267583d793ef3153e88760ae5aaebe751501e2cde7f1d7ba9530324dd24`.
+
+## Frozen design
+
+After verified publication of 0.1.14 at
+`08319af02eeee95b88978dd126e392fec67719de`, eight composite investigation tasks
+were frozen across pinned FastAPI, GitHub CLI, Requests and Phoenix sources.
+Each requested an implementation, a cross-file caller, an existing regression
+test and a justified negative control. Independent source review bound 73
+anchors in 22 files before client outcomes. The I07 primary investigation was
+new, but one supporting caller/function anchor had been consumed previously;
+that overlap was disclosed before launch. Supporting spans are not all fresh.
+
+Both arms used Claude Code 2.1.287 and the observed model identity
+`claude-sonnet-5-5`, with native Read/Grep/Glob and structured serialization.
+The optional Attocode arm additionally advertised the 16 daily MCP tools from
+the released source; using them was not compulsory. Inputs, client, dependencies,
+operators and source snapshots were frozen. Prompts and labels for unlaunched
+tasks are not exported.
+
+Limits were two workers, 180 seconds end to end per attempt, a shared 50-minute
+window, 2 GiB managed disk and aggregate RSS, at most 16 scored requests and two
+excluded wiring requests. Admission reserved $0.45 per request within the $10
+pilot allowance. Retries and replacements after an outcome were forbidden.
+Expansion required at least two additional independently fully correct answers
+and aggregate reported-cost and end-to-end time ratios each at most 1.25.
+
+## Observed execution
+
+Both excluded wiring checks passed. Each observed the expected client/tool
+inventory, denied the owned outside-root Read canary without returning its
+contents, and produced a valid synthetic answer. These checks are excluded from
+quality and paired cost/time comparisons.
+
+The first scored pair was I07, a Phoenix cache-policy investigation:
+
+| Attempt | Observed outcome | End-to-end observation | Reported cost |
+| --- | --- | --- | --- |
+| Attocode available | Incomplete: first structured-output call malformed; corrected final answer does not override protocol failure | 35.517 s | $0.081174 |
+| Native | Incomplete: malformed output and four unavailable pseudo-tool calls; interrupted when the other worker failed | 35.745 s to cancellation | Unknown |
+
+The Attocode-available client made six native read/search calls and **zero MCP
+calls**. Its first serialization call supplied unexpected root properties and
+string values where nested role objects were required. The client rejected it,
+then accepted a corrected serialization. The final answer passed schema and
+exact-source-quote checks. The frozen auditor nevertheless validates every
+`StructuredOutput` call and rejects the earlier malformed arguments, so the
+worker returned incomplete. Final schema/quote validity is not
+a semantic correctness grade.
+
+The native trace also contains a malformed serialization call. It then invoked
+unavailable tools named `caller`, `regression_test`, `negative_control` and
+`uncertainties`; the client explicitly rejected all four. The controller stopped
+the active pair after the first worker failure. Native has no terminal result
+or reported total cost. A low-level supervision status does not establish study
+completion. No additional client call, retry or replacement followed.
+
+These traces establish a client/model serialization and study-protocol failure
+boundary. They do not establish its underlying cause or a server defect. In
+particular, the Attocode-available run did not exercise an MCP operation.
+
+## Decision, accounting and evidence limits
+
+The frozen summary is incomplete. Quality delta and both paired ratios remain
+null. No semantic grades were assigned: one mechanically valid final answer
+cannot satisfy an eight-pair study. Only I07 was exposed to scored clients; the
+other seven tasks were not run and are not replacement attempts for this study.
+
+Known reported costs, including both excluded wiring checks, subtotal
+**$0.1823414**. This is partial accounting; the interrupted native cost is unknown,
+so total cost is unknown. Subscription extra usage was off at preflight.
+API-equivalent reported cost is not an invoice. No further pilot spending is
+authorized by unused reservations. Remote performance allocation earlier in the
+plan started no worker; its estimated compute cost remains $0, invoices unchecked.
+
+The scored controller sampled peak aggregate RSS of approximately 731 MiB and
+managed storage of 86.8 MiB. Owned temporary repositories were removed; the
+independent audit found no owned pilot process. Frozen source/input/operator and
+client identities match. Both scored failures occurred before the final
+temporary-source hash checks, so their unchanged temporary-copy identity is not
+claimed. Read-only post-run audit imports accidentally created two bytecode files;
+those additions were preserved separately, and original frozen identities were
+restored and verified. Operator source and raw attempts were unchanged.
+
+Raw events, failed attempts, initial preparation/export errors, subscription
+evidence and private answer mappings remain local. The archive exposes structural
+tool-call summaries, accounting, resource observations and immutable evidence
+hashes; it excludes raw answers, private paths and unlaunched prompts/labels.
+After extracting it, `python replay.py` verifies archive member hashes and
+recomputes the incomplete accounting/protocol observations. This is not a replay
+of semantic grading or a new model run.
+
+The approved local-first batch closes with 0.1.14 published, the restart shortcut
+withheld without a matched measurement, the callback-span candidate rejected on
+consumed development, and this pilot incomplete. Native cold-parse and Rust
+readiness causes, restart cost and first-page relevance remain open. No default
+ranking, profile, embedding or model policy changes follow from this pilot.