chore(ci): remove Dependabot, add Claude + Copilot code review

.github/workflows/ci.yml

@@ -26,9 +26,11 @@ jobs:
           cmake -B build -G Ninja \
             -DCMAKE_C_COMPILER=gcc-12 \
             -DCMAKE_CXX_COMPILER=g++-12 \
-            -DBUILD_TESTS=ON
+            -DEAI_BUILD_TESTS=ON
           cmake --build build --parallel $(nproc)
-          cd build && ctest --output-on-failure || true
+          # No `|| true`: a failing test must fail the job. --no-tests=error
+          # catches an empty test set, which ctest otherwise reports as success.
+          cd build && ctest --output-on-failure --no-tests=error
 
   # ── Python Tests ──────────────────────────────────────────────────────────
   test-python:

.github/workflows/claude-code-review.yml

@@ -0,0 +1,132 @@
+# Claude code review.
+#
+# This complements — it does not replace — the automatic Copilot review that
+# runs on every PR via the "Code Quality Copilot review for default branch"
+# ruleset. Copilot does the fast first pass; Claude does the deeper one, and
+# only when asked, so routine PRs cost nothing.
+#
+# Two ways in:
+#   1. Mention @claude in an issue, a PR comment, or a review comment.
+#   2. Add the `deep-review` label to a PR for a full review pass.
+#
+# Requires an ANTHROPIC_API_KEY secret (org-level is easiest — one secret
+# covers every repo). Without it both jobs fail fast with a clear message
+# rather than reviewing silently with no credentials.
+name: Claude Code Review
+
+on:
+  issue_comment:
+    types: [created]
+  pull_request_review_comment:
+    types: [created]
+  pull_request_review:
+    types: [submitted]
+  issues:
+    types: [opened, assigned, labeled]
+  pull_request:
+    types: [labeled]
+
+# Never run two reviews on the same PR at once; a new trigger supersedes the
+# one in flight.
+concurrency:
+  group: claude-review-${{ github.event.pull_request.number || github.event.issue.number || github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  # ── 1. On-demand: someone wrote @claude ──────────────────────────────────
+  mention:
+    if: |
+      github.event_name != 'pull_request' && (
+        (github.event_name == 'issue_comment' && contains(github.event.comment.body, '@claude')) ||
+        (github.event_name == 'pull_request_review_comment' && contains(github.event.comment.body, '@claude')) ||
+        (github.event_name == 'pull_request_review' && contains(github.event.review.body, '@claude')) ||
+        (github.event_name == 'issues' && (contains(github.event.issue.body, '@claude') || contains(github.event.issue.title, '@claude')))
+      )
+    runs-on: ubuntu-latest
+    timeout-minutes: 20
+    permissions:
+      contents: read
+      pull-requests: write
+      issues: write
+      id-token: write
+    steps:
+      - name: Check credentials
+        env:
+          KEY: ${{ secrets.ANTHROPIC_API_KEY }}
+        run: |
+          if [ -z "$KEY" ]; then
+            echo "::error::ANTHROPIC_API_KEY is not set. Add it as an organisation or repository secret."
+            exit 1
+          fi
+
+      - uses: actions/checkout@v6
+        with:
+          fetch-depth: 1
+
+      - uses: anthropics/claude-code-action@v1
+        with:
+          anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
+          use_sticky_comment: true
+
+  # ── 2. Full pass: PR labelled `deep-review` ──────────────────────────────
+  deep-review:
+    if: github.event_name == 'pull_request' && github.event.label.name == 'deep-review'
+    runs-on: ubuntu-latest
+    timeout-minutes: 30
+    permissions:
+      contents: read
+      pull-requests: write
+      id-token: write
+    steps:
+      - name: Check credentials
+        env:
+          KEY: ${{ secrets.ANTHROPIC_API_KEY }}
+        run: |
+          if [ -z "$KEY" ]; then
+            echo "::error::ANTHROPIC_API_KEY is not set. Add it as an organisation or repository secret."
+            exit 1
+          fi
+
+      - uses: actions/checkout@v6
+        with:
+          fetch-depth: 1
+
+      - uses: anthropics/claude-code-action@v1
+        with:
+          anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
+          use_sticky_comment: true
+          prompt: |
+            REPO: ${{ github.repository }}
+            PR NUMBER: ${{ github.event.pull_request.number }}
+
+            Review this pull request. The PR branch is already checked out in
+            the working directory.
+
+            This repository follows the standards in its own CLAUDE.md,
+            QUALITY.md, TESTING.md, SECURITY.md and VERIFY.md — read the ones
+            that apply to the changed files and review against them, not
+            against generic style preferences.
+
+            Prioritise, in this order:
+              1. Correctness — logic errors, unhandled failure paths, integer
+                 overflow/underflow, memory safety, race conditions.
+              2. Security — unvalidated input crossing a boundary, injection,
+                 authorization that checks only authentication, secrets in
+                 source or logs.
+              3. Tests — does a new behaviour have a test that would actually
+                 fail without the fix? Flag assertions that cannot fail, and
+                 tests that only check the negative case.
+              4. Maintainability — only where it genuinely impedes a reader.
+
+            Report what you verified and what you could not. Say plainly when
+            a concern is unverified rather than implying you checked it. If
+            nothing needs changing, say so in one line — do not manufacture
+            findings.
+
+            Use `gh pr comment` for top-level feedback.
+            Use `mcp__github_inline_comment__create_inline_comment` (with
+            `confirmed: true`) for specific lines.
+            Only post GitHub comments — do not return review text as a message.
+
+          claude_args: |
+            --allowedTools "mcp__github_inline_comment__create_inline_comment,Bash(gh pr comment:*),Bash(gh pr diff:*),Bash(gh pr view:*),Bash(gh pr checks:*)"

TASKS.md

@@ -17,7 +17,9 @@ Status is one of: `todo`, `in-progress`, `blocked`, `review`, `done`.
 
 | ID | Task | Owner | Verified by | Evidence |
 |----|------|-------|-------------|----------|
-| —  | None yet. | — | — | — |
+| T-001 | CI ran zero tests and reported success | testing | reviewer | `ci.yml` configured with `-DBUILD_TESTS=ON`, but this project's option is `EAI_BUILD_TESTS`, so no test binary was built; `ctest` exits 0 on an empty test set (verified: exit code 0). The step also ended in `|| true`, which would have discarded a real failure as well. Fixed the flag, removed `|| true`, added `--no-tests=error`. CI now builds and runs 24 tests, all passing. |
+| T-002 | Fix an integer underflow in the GGUF model-file parser | security | reviewer | `formats/src/gguf_loader.c` rounded the data-section offset up to a 32-byte boundary, which on a truncated file lands past EOF. `file_end - data_start` was then negative and cast to `size_t`, so `ctx->data_size` became 0xffffffffffffffe4 (~1.8e19). That value reached `malloc()` — ASan: `allocation-size-too-big` — and, because the `if (ctx->data)` guard left the field set, stayed in `ctx->data_size` for every later consumer to read as a length. Model files are attacker-supplied input. Now bounds-checked against EOF, with `data_size` set from the actual `fread` return so a short read cannot report more bytes than were read. `eai_format_tests` and `eai_format_ext_tests` pass under ASan/UBSan. |
+| T-003 | Fix a memory leak in the adaptive-profile test | testing | reviewer | `tests/test_adaptive.c:305` loaded the `adaptive-edge` profile, which `strdup()`s seven tool names plus a provider string, and never called `eai_config_free()`. LeakSanitizer reported the direct leaks from `config.c:153` and `:158`. 24/24 eAI tests now pass under `-fsanitize=address,undefined`. |
 
 ---
 

formats/src/gguf_loader.c

@@ -6,6 +6,7 @@
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
+#include <limits.h>
 
 #define LOG_MOD "gguf"
 
@@ -134,25 +135,48 @@ eai_status_t eai_gguf_load(const char *path, gguf_context_t *ctx)
         }
     }
 
-    /* Data section starts after alignment to 32 bytes */
-    long pos = ftell(fp);
-    long aligned_pos = (pos + 31) & ~31L;
-    fseek(fp, aligned_pos, SEEK_SET);
+    /* Position reached after the header/KV/tensor-info sections. */
+    long ftell_after_metadata = ftell(fp);
 
-    /* Read tensor data */
-    long data_start = ftell(fp);
-    fseek(fp, 0, SEEK_END);
+    /* Establish the true end of file first, so the data section can be
+     * bounds-checked against it. */
+    if (fseek(fp, 0, SEEK_END) != 0) { fclose(fp); eai_gguf_free(ctx); return EAI_ERR_IO; }
     long file_end = ftell(fp);
+    if (file_end < 0) { fclose(fp); eai_gguf_free(ctx); return EAI_ERR_IO; }
+
+    /* Data section starts after alignment to 32 bytes */
+    long pos = ftell_after_metadata;
+    if (pos < 0) { fclose(fp); eai_gguf_free(ctx); return EAI_ERR_IO; }
+    long aligned_pos = (pos > LONG_MAX - 31) ? file_end : ((pos + 31) & ~31L);
+
+    /* Rounding up to the alignment boundary can land past the end of a
+     * truncated file. Computing file_end - data_start unguarded then yielded a
+     * negative value that became ~1.8e19 when cast to size_t, and that value
+     * reached malloc() and stayed in ctx->data_size for every later consumer to
+     * read as a length. Treat "starts at or past EOF" as an empty data section. */
+    if (aligned_pos >= file_end) {
+        ctx->data = NULL;
+        ctx->data_size = 0;
+        fclose(fp);
+        return EAI_OK;
+    }
+
+    long data_start = aligned_pos;
     ctx->data_size = (size_t)(file_end - data_start);
 
-    if (ctx->data_size > 0) {
-        ctx->data = (uint8_t *)malloc(ctx->data_size);
-        if (ctx->data) {
-            fseek(fp, data_start, SEEK_SET);
-            fread(ctx->data, 1, ctx->data_size, fp);
-        }
+    ctx->data = (uint8_t *)malloc(ctx->data_size);
+    if (!ctx->data) { ctx->data_size = 0; fclose(fp); eai_gguf_free(ctx); return EAI_ERR_NOMEM; }
+
+    if (fseek(fp, data_start, SEEK_SET) != 0) {
+        free(ctx->data); ctx->data = NULL; ctx->data_size = 0;
+        fclose(fp); eai_gguf_free(ctx); return EAI_ERR_IO;
     }
 
+    /* Keep data_size consistent with what was actually read, so a short read
+     * cannot leave a length longer than the buffer's valid contents. */
+    size_t got = fread(ctx->data, 1, ctx->data_size, fp);
+    ctx->data_size = got;
+
     fclose(fp);
     return EAI_OK;
 }

tests/test_adaptive.c

@@ -307,6 +307,8 @@ TEST(test_adaptive_profile_loading)
     assert(cfg.adaptive.lora_rank == 8);
     assert(cfg.adaptive.max_training_memory_mb == 512);
     assert(cfg.tool_count == 7);
+    /* The profile strdup()s each tool name; without this the test leaks them. */
+    eai_config_free(&cfg);
     return 0;
 }