cctx-cli 1.2.0__tar.gz → 1.4.0__tar.gz

cctx_cli-1.4.0/CHANGELOG.md

@@ -0,0 +1,230 @@
+# CHANGELOG
+
+<!-- version list -->
+
+## v1.4.0 (2026-05-20)
+
+### Bug Fixes
+
+- Deduplicate harvest check import, align severity badge output
+  ([#87](https://github.com/jacquardlabs/cctx/pull/87),
+  [`ee08734`](https://github.com/jacquardlabs/cctx/commit/ee0873431383b285769195efc4b2f70f5d07cdeb))
+
+- Move defaultdict import to top-level, add _words() return type
+  ([#87](https://github.com/jacquardlabs/cctx/pull/87),
+  [`ee08734`](https://github.com/jacquardlabs/cctx/commit/ee0873431383b285769195efc4b2f70f5d07cdeb))
+
+- Use removeprefix instead of lstrip to preserve .claude/skills/ dot prefix
+  ([#87](https://github.com/jacquardlabs/cctx/pull/87),
+  [`ee08734`](https://github.com/jacquardlabs/cctx/commit/ee0873431383b285769195efc4b2f70f5d07cdeb))
+
+### Documentation
+
+- M15 harvest --check depth design spec ([#87](https://github.com/jacquardlabs/cctx/pull/87),
+  [`ee08734`](https://github.com/jacquardlabs/cctx/commit/ee0873431383b285769195efc4b2f70f5d07cdeb))
+
+- M15 harvest --check depth implementation plan
+  ([#87](https://github.com/jacquardlabs/cctx/pull/87),
+  [`ee08734`](https://github.com/jacquardlabs/cctx/commit/ee0873431383b285769195efc4b2f70f5d07cdeb))
+
+### Features
+
+- --check-severity flag and severity badges in harvest --check output
+  ([#87](https://github.com/jacquardlabs/cctx/pull/87),
+  [`ee08734`](https://github.com/jacquardlabs/cctx/commit/ee0873431383b285769195efc4b2f70f5d07cdeb))
+
+- Check_contradictions() — always/never keyword heuristic
+  ([#87](https://github.com/jacquardlabs/cctx/pull/87),
+  [`ee08734`](https://github.com/jacquardlabs/cctx/commit/ee0873431383b285769195efc4b2f70f5d07cdeb))
+
+- Check_redundancy() — Jaccard similarity ≥ 0.8 on section word sets
+  ([#87](https://github.com/jacquardlabs/cctx/pull/87),
+  [`ee08734`](https://github.com/jacquardlabs/cctx/commit/ee0873431383b285769195efc4b2f70f5d07cdeb))
+
+- Check_staleness() — backtick function refs grepped against project source
+  ([#87](https://github.com/jacquardlabs/cctx/pull/87),
+  [`ee08734`](https://github.com/jacquardlabs/cctx/commit/ee0873431383b285769195efc4b2f70f5d07cdeb))
+
+- CheckSeverity enum, severity field on CheckFinding, new CheckIssue values
+  ([#87](https://github.com/jacquardlabs/cctx/pull/87),
+  [`ee08734`](https://github.com/jacquardlabs/cctx/commit/ee0873431383b285769195efc4b2f70f5d07cdeb))
+
+- Harvest --check depth — contradiction, redundancy, staleness detectors + --check-severity
+  ([#87](https://github.com/jacquardlabs/cctx/pull/87),
+  [`ee08734`](https://github.com/jacquardlabs/cctx/commit/ee0873431383b285769195efc4b2f70f5d07cdeb))
+
+- Wire all four detectors into check_claude_md ([#87](https://github.com/jacquardlabs/cctx/pull/87),
+  [`ee08734`](https://github.com/jacquardlabs/cctx/commit/ee0873431383b285769195efc4b2f70f5d07cdeb))
+
+### Refactoring
+
+- Check_redundancy — compute _words once per section, remove dead union guard
+  ([#87](https://github.com/jacquardlabs/cctx/pull/87),
+  [`ee08734`](https://github.com/jacquardlabs/cctx/commit/ee0873431383b285769195efc4b2f70f5d07cdeb))
+
+- Check_staleness — module-level _STALENESS_EXCLUDED, min-len in regex, per-file search
+  ([#87](https://github.com/jacquardlabs/cctx/pull/87),
+  [`ee08734`](https://github.com/jacquardlabs/cctx/commit/ee0873431383b285769195efc4b2f70f5d07cdeb))
+
+
+## v1.3.0 (2026-05-17)
+
+### Bug Fixes
+
+- Drop unused turn_number from result_map in _find_pairs
+  ([#86](https://github.com/jacquardlabs/cctx/pull/86),
+  [`cefc438`](https://github.com/jacquardlabs/cctx/commit/cefc438f9ff638ba2abf529663b5e24707f03bbb))
+
+- Restore WHY comment, fix_key != failure_key guard, tighten tuple annotation
+  ([#86](https://github.com/jacquardlabs/cctx/pull/86),
+  [`cefc438`](https://github.com/jacquardlabs/cctx/commit/cefc438f9ff638ba2abf529663b5e24707f03bbb))
+
+- Ruff lint failures (E501, F401, E741, I001) ([#86](https://github.com/jacquardlabs/cctx/pull/86),
+  [`cefc438`](https://github.com/jacquardlabs/cctx/commit/cefc438f9ff638ba2abf529663b5e24707f03bbb))
+
+### Documentation
+
+- M14 project-pattern-detection implementation plan
+  ([#86](https://github.com/jacquardlabs/cctx/pull/86),
+  [`cefc438`](https://github.com/jacquardlabs/cctx/commit/cefc438f9ff638ba2abf529663b5e24707f03bbb))
+
+- M14 project-specific pattern detection design spec
+  ([#86](https://github.com/jacquardlabs/cctx/pull/86),
+  [`cefc438`](https://github.com/jacquardlabs/cctx/commit/cefc438f9ff638ba2abf529663b5e24707f03bbb))
+
+- Note why harvest --since skips project_specific.detect()
+  ([#86](https://github.com/jacquardlabs/cctx/pull/86),
+  [`cefc438`](https://github.com/jacquardlabs/cctx/commit/cefc438f9ff638ba2abf529663b5e24707f03bbb))
+
+### Features
+
+- Add ProjectPattern model, AggregateReport.project_patterns, FindingKind.PROJECT_PATTERN
+  ([#86](https://github.com/jacquardlabs/cctx/pull/86),
+  [`cefc438`](https://github.com/jacquardlabs/cctx/commit/cefc438f9ff638ba2abf529663b5e24707f03bbb))
+
+- Aggregate.run() returns (Diagnosis, SessionTrace) pairs
+  ([#86](https://github.com/jacquardlabs/cctx/pull/86),
+  [`cefc438`](https://github.com/jacquardlabs/cctx/commit/cefc438f9ff638ba2abf529663b5e24707f03bbb))
+
+- Generate_from_patterns() — CLAUDE.md patches from ProjectPatterns
+  ([#86](https://github.com/jacquardlabs/cctx/pull/86),
+  [`cefc438`](https://github.com/jacquardlabs/cctx/commit/cefc438f9ff638ba2abf529663b5e24707f03bbb))
+
+- M14 project-specific pattern detection ([#86](https://github.com/jacquardlabs/cctx/pull/86),
+  [`cefc438`](https://github.com/jacquardlabs/cctx/commit/cefc438f9ff638ba2abf529663b5e24707f03bbb))
+
+- Project_specific.detect() — cross-session failure/fix pattern detector
+  ([#86](https://github.com/jacquardlabs/cctx/pull/86),
+  [`cefc438`](https://github.com/jacquardlabs/cctx/commit/cefc438f9ff638ba2abf529663b5e24707f03bbb))
+
+- Render_aggregate() shows project-specific patterns table
+  ([#86](https://github.com/jacquardlabs/cctx/pull/86),
+  [`cefc438`](https://github.com/jacquardlabs/cctx/commit/cefc438f9ff638ba2abf529663b5e24707f03bbb))
+
+- Wire project_specific.detect() into autopsy and harvest --since paths
+  ([#86](https://github.com/jacquardlabs/cctx/pull/86),
+  [`cefc438`](https://github.com/jacquardlabs/cctx/commit/cefc438f9ff638ba2abf529663b5e24707f03bbb))
+
+
+## v1.2.0 (2026-05-17)
+
+### Features
+
+- --until DATE, autopsy --json, export --format json (M12 #77 #78 #79)
+  ([#84](https://github.com/jacquardlabs/cctx/pull/84),
+  [`803b5f1`](https://github.com/jacquardlabs/cctx/commit/803b5f190404679ddef4cbbec7478d04c57b8413))
+
+
+## v1.1.0 (2026-05-17)
+
+### Chores
+
+- Add skip-existing to pypi publish action
+  ([`23d7e16`](https://github.com/jacquardlabs/cctx/commit/23d7e16e18074da3c25899ba98298100ad3c1ad3))
+
+### Features
+
+- M9 polish — verdict headline, --top N, --turn N
+  ([#83](https://github.com/jacquardlabs/cctx/pull/83),
+  [`b0d2f27`](https://github.com/jacquardlabs/cctx/commit/b0d2f273a373c5a2f52c9de3a3fb2721da59c4f5))
+
+- M9 polish — verdict headline, --top N, and --turn N
+  ([#83](https://github.com/jacquardlabs/cctx/pull/83),
+  [`b0d2f27`](https://github.com/jacquardlabs/cctx/commit/b0d2f273a373c5a2f52c9de3a3fb2721da59c4f5))
+
+### Refactoring
+
+- Cache verdict, fix markup=False bug, use reverse=True
+  ([#83](https://github.com/jacquardlabs/cctx/pull/83),
+  [`b0d2f27`](https://github.com/jacquardlabs/cctx/commit/b0d2f273a373c5a2f52c9de3a3fb2721da59c4f5))
+
+
+## v1.0.0 (2026-05-17)
+
+### Continuous Integration
+
+- Add python-semantic-release for fully automated CD
+  ([`9844921`](https://github.com/jacquardlabs/cctx/commit/98449213e5b3bd597c47d54e4d5043e245adafe4))
+
+- Add workflow_dispatch to release.yml for manual trigger
+  ([`08ac9f8`](https://github.com/jacquardlabs/cctx/commit/08ac9f80eb6ceee7b57155852febc4274cbaf3b0))
+
+
+## v0.2.0 (2026-05-16)
+
+### Bug Fixes
+
+- Ruff lint — B904, E402, E501, F841 across cli, tests, and renderers
+  ([`fa7105f`](https://github.com/jacquardlabs/cctx/commit/fa7105fef340d89136f1996b86824e30d080a730))
+
+- Trace TUI token sum within line-length limit
+  ([`5b7416c`](https://github.com/jacquardlabs/cctx/commit/5b7416c5b92790cfe66f1e53f20891ecaf6e03b0))
+
+### Chores
+
+- Bump version to 0.2.0, update PRODUCT.md and CLAUDE.md
+  ([`828ed49`](https://github.com/jacquardlabs/cctx/commit/828ed4997df9f4a264669bc38f4b10588a151f1c))
+
+### Documentation
+
+- Add CI usage section clarifying harvest is local-only
+  ([`c526408`](https://github.com/jacquardlabs/cctx/commit/c526408a35749858e1c0b0b6ba42aea95bb8f621))
+
+### Features
+
+- **#64,#63**: Tool-thrash and dead-end exploration classifiers
+  ([`14f8f45`](https://github.com/jacquardlabs/cctx/commit/14f8f45f9d3f4ef2fefac374d3e4cea36185c60d))
+
+- **#65**: Harvest v2 — route patches to any .md target (.claude/rules/, .claude/skills/)
+  ([`06ef9b7`](https://github.com/jacquardlabs/cctx/commit/06ef9b7a8ae9abacea54c2c826efc3fb6e6e80be))
+
+- **#66**: Cctx harvest --check audits CLAUDE.md for dead refs and empty sections
+  ([`a3be1d0`](https://github.com/jacquardlabs/cctx/commit/a3be1d0923d387b6830b10c7c2c5acf34a3b8917))
+
+- **#67**: Interactive aggregate drill-down; --check docs in README
+  ([`3db5429`](https://github.com/jacquardlabs/cctx/commit/3db5429dede40777b23d83e32a4b15a8c0e82a16))
+
+- **#68**: --since accepts 7d, 2w, YYYY-MM-DD, and date ranges
+  ([`434d7c4`](https://github.com/jacquardlabs/cctx/commit/434d7c448406e0a5465380ed395e1fadaa1c0db1))
+
+- **#69**: Annotate costs as estimates (~85–95%) in terminal and HTML output
+  ([`5a49889`](https://github.com/jacquardlabs/cctx/commit/5a49889bd43555d30a58c2b78eddf6acbb0d8e97))
+
+- **#70**: Cctx watch — live waste signals during an active session
+  ([`f533a13`](https://github.com/jacquardlabs/cctx/commit/f533a13e1f087da00457ce7b8215934f20c403a2))
+
+- **#72**: Cctx autopsy --github-summary writes findings to GitHub Actions job summary
+  ([`df91256`](https://github.com/jacquardlabs/cctx/commit/df91256b58b75a9bbe9594a000b00dee2ac2fbc5))
+
+- **#73**: Cctx GitHub Action (composite) + --fail-on-findings flag
+  ([`9229584`](https://github.com/jacquardlabs/cctx/commit/9229584ea192ce1b1ee6718721d981cba0ca13e0))
+
+### Refactoring
+
+- Consolidate KIND_LABEL, fix private import, clean up watcher tests
+  ([`31ad4d7`](https://github.com/jacquardlabs/cctx/commit/31ad4d75885eed9c8238c4eb1c08bd5d1ba51a15))
+
+
+## v0.1.0 (2026-05-16)
+
+- Initial Release

{cctx_cli-1.2.0 → cctx_cli-1.4.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: cctx-cli
-Version: 1.2.0
+Version: 1.4.0
 Summary: Diagnose Claude Code sessions — find what went wrong, what it cost, and what to add to CLAUDE.md
 Author: Jacquard Labs
 License-Expression: MIT

{cctx_cli-1.2.0 → cctx_cli-1.4.0}/cctx/__init__.py

@@ -1,3 +1,3 @@
 """cctx: profile, debug, and optimize Claude Code and Agent SDK sessions."""
 
-__version__ = "1.2.0"
+__version__ = "1.4.0"

{cctx_cli-1.2.0 → cctx_cli-1.4.0}/cctx/cli.py

@@ -20,6 +20,7 @@ import rich_click as click
 
 from cctx import diagnostician
 from cctx.diagnostician import aggregate
+from cctx.diagnostician.patterns import project_specific
 from cctx.discovery import complete_project as _complete_project
 from cctx.models import KIND_LABEL, AggregateReport
 from cctx.parsers.claude_code import parse_session
@@ -146,22 +147,32 @@ def _render_check_findings(findings: list, target_dir: Path) -> None:
     from rich.console import Console
     from rich.rule import Rule
 
+    from cctx.harvest import CheckIssue, CheckSeverity
+
     con = Console()
     claude_md_path = target_dir / "CLAUDE.md"
     con.print(Rule(f"cctx harvest --check — {claude_md_path}"))
     if not findings:
-        con.print("✓ CLAUDE.md looks clean — no dead references or empty sections.")
+        con.print("✓ CLAUDE.md looks clean — no issues found.")
         return
     con.print(f"{len(findings)} issue(s) found:\n")
-    from cctx.harvest import CheckIssue
     _ISSUE_LABEL = {
-        CheckIssue.DEAD_FILE_REF:  "dead file reference",
-        CheckIssue.DEAD_SKILL_REF: "dead skill reference",
-        CheckIssue.EMPTY_SECTION:  "empty section",
+        CheckIssue.DEAD_FILE_REF:    "dead file reference",
+        CheckIssue.DEAD_SKILL_REF:   "dead skill reference",
+        CheckIssue.EMPTY_SECTION:    "empty section",
+        CheckIssue.CONTRADICTION:    "contradiction",
+        CheckIssue.REDUNDANCY:       "redundancy",
+        CheckIssue.STALE_IDENTIFIER: "stale identifier",
+    }
+    _SEV_BADGE = {
+        CheckSeverity.HIGH:   "[HIGH]",
+        CheckSeverity.MEDIUM: "[MED]",
+        CheckSeverity.LOW:    "[LOW]",
     }
     for f in findings:
+        badge = _SEV_BADGE.get(f.severity, "      ")
         label = _ISSUE_LABEL.get(f.issue, f.issue.value)
-        con.print(f"  [{f.heading}]  {label}: {f.detail}")
+        con.print(f"  {badge:<6}  {f.heading}  {label}: {f.detail}")
 
 
 @click.group()
@@ -358,11 +369,14 @@ def autopsy(
                     f"Invalid --until date '{until_date}'. Expected YYYY-MM-DD."
                 ) from None
             label = f"{label} until {until_date.strip()}"
-        diagnoses = aggregate.run(project_dir, start, end)
+        pairs = aggregate.run(project_dir, start, end)
+        diagnoses = [d for d, _ in pairs]
         ev = evidence_mod.accumulate(diagnoses)
         if top_n is not None:
             ev = dict(sorted(ev.items(), key=lambda x: x[1].session_count, reverse=True)[:top_n])
-        patches = claude_md.generate_from_evidence(ev)
+        patterns = project_specific.detect(pairs)
+        pattern_patches = claude_md.generate_from_patterns(patterns)
+        patches = claude_md.generate_from_evidence(ev) + pattern_patches
         report = AggregateReport(
             period_label=label,
             sessions_analysed=len(diagnoses),
@@ -371,6 +385,7 @@ def autopsy(
             waste_cost_usd=sum(d.waste_cost_usd for d in diagnoses),
             by_kind=ev,
             patches=patches,
+            project_patterns=patterns,
         )
         render_aggregate(report)
         _aggregate_drilldown(report, diagnoses)
@@ -544,6 +559,14 @@ def trace(target: Path | None, latest: bool) -> None:
     default=False,
     help="Audit existing CLAUDE.md for dead references and empty sections. Exit 1 if findings.",
 )
+@click.option(
+    "--check-severity",
+    "check_severity",
+    default="MEDIUM",
+    type=click.Choice(["LOW", "MEDIUM", "HIGH"], case_sensitive=False),
+    show_default=True,
+    help="Minimum severity that causes --check to exit 1.",
+)
 def harvest(
     target: Path,
     since: str | None,
@@ -551,15 +574,27 @@ def harvest(
     dry_run: bool,
     target_dir: Path | None,
     check_mode: bool,
+    check_severity: str,
 ) -> None:
     """Apply autopsy patches to CLAUDE.md."""
     from cctx.harvest import apply_patches, check_claude_md, preview_patches
 
     if check_mode:
+        from cctx.harvest import CheckSeverity
         resolved_dir = target_dir or Path.cwd()
         findings = check_claude_md(resolved_dir)
         _render_check_findings(findings, resolved_dir)
-        raise SystemExit(1 if findings else 0)
+        _SEVERITY_ORDER = {
+            CheckSeverity.LOW: 0,
+            CheckSeverity.MEDIUM: 1,
+            CheckSeverity.HIGH: 2,
+        }
+        threshold = CheckSeverity(check_severity.lower())
+        triggering = [
+            f for f in findings
+            if _SEVERITY_ORDER[f.severity] >= _SEVERITY_ORDER[threshold]
+        ]
+        raise SystemExit(1 if triggering else 0)
 
     if apply_mode and dry_run:
         raise click.UsageError("--apply and --dry-run are mutually exclusive.")
@@ -571,8 +606,11 @@ def harvest(
     if since is not None:
         project_dir = target if target.is_dir() else target.parent
         start, end, _label = parse_since(since)
-        diagnoses = aggregate.run(project_dir, start, end)
+        pairs = aggregate.run(project_dir, start, end)
+        diagnoses = [d for d, _ in pairs]
         ev = evidence_mod.accumulate(diagnoses)
+        # project_specific.detect() intentionally omitted: pattern patches need human review
+        # (autopsy shows them; harvest doesn't auto-apply).
         patches = claude_md.generate_from_evidence(ev)
     else:
         if target.is_dir():

{cctx_cli-1.2.0 → cctx_cli-1.4.0}/cctx/diagnostician/aggregate.py

@@ -1,10 +1,11 @@
 """Cross-session aggregator.
 
-run(project_dir, start, end) -> list[Diagnosis]
+run(project_dir, start, end) -> list[tuple[Diagnosis, SessionTrace]]
 
 Discovers session JSONL files in project_dir modified within [start, end],
-parses each one, runs the per-session diagnostician, and returns the list of
-Diagnoses. The CLI orchestrates the recommender call separately.
+parses each one, runs the per-session diagnostician, and returns
+(Diagnosis, SessionTrace) pairs. The CLI orchestrates recommender and
+project-specific detection separately.
 """
 from __future__ import annotations
 
@@ -17,12 +18,14 @@ from cctx.parsers.claude_code import parse_session
 from cctx.tokenizer import tokenize_session
 
 if TYPE_CHECKING:
-    from cctx.models import Diagnosis
+    from cctx.models import Diagnosis, SessionTrace
 
 UTC = timezone.utc
 
 
-def run(project_dir: Path, start: datetime, end: datetime) -> list[Diagnosis]:
+def run(
+    project_dir: Path, start: datetime, end: datetime
+) -> list[tuple[Diagnosis, SessionTrace]]:
     paths = sorted(project_dir.glob("*.jsonl"), key=lambda p: p.stat().st_mtime)
 
     result = []
@@ -33,7 +36,7 @@ def run(project_dir: Path, start: datetime, end: datetime) -> list[Diagnosis]:
         try:
             trace = tokenize_session(parse_session(path))
             diagnosis = diagnostician.run(trace)
-            result.append(diagnosis)
+            result.append((diagnosis, trace))
         except Exception:
             continue  # skip corrupt sessions; don't fail the whole run
     return result

cctx_cli-1.4.0/cctx/diagnostician/patterns/project_specific.py

@@ -0,0 +1,179 @@
+"""Project-specific pattern detector.
+
+detect(pairs) -> list[ProjectPattern]
+
+Finds (tool_name, failure_key, fix_key) triples that recur in 3+ sessions.
+Bash normalization uses first 3 tokens for cross-session fuzzy matching
+(intentionally looser than retry_loop). No LLM calls.
+"""
+from __future__ import annotations
+
+import json
+from collections import defaultdict
+from typing import TYPE_CHECKING
+
+from cctx.models import ProjectPattern
+from cctx.pricing import price_per_tok
+
+if TYPE_CHECKING:
+    from cctx.models import Diagnosis, SessionTrace, ToolResult
+
+MIN_SESSIONS = 3
+FIX_WINDOW = 10  # turns after last failure to search for the fix
+
+
+def _normalize_key(tool_name: str, tool_input: dict) -> str:
+    match tool_name:
+        case "Bash":
+            tokens = tool_input.get("command", "").strip().split()
+            return " ".join(tokens[:3])
+        case "Edit" | "Read" | "Write":
+            return tool_input.get("file_path", "")
+        case "Grep" | "Glob":
+            return tool_input.get("pattern", "")
+        case _:
+            return json.dumps(tool_input, sort_keys=True)
+
+
+def _is_error(result: ToolResult) -> bool:
+    if result.is_error:
+        return True
+    c = result.content
+    return c.startswith("Error:") or c.startswith("error:") or c.startswith("FAILED")
+
+
+def _find_pairs(trace: SessionTrace) -> list[dict]:
+    """Find failure/fix pairs within one session."""
+    result_map: dict[str, ToolResult] = {}
+    for turn in trace.turns:
+        for tr in turn.tool_results:
+            result_map[tr.tool_use_id] = tr
+
+    records = []
+    for turn in trace.turns:
+        if turn.role != "assistant":
+            continue
+        for tu in turn.tool_uses:
+            result = result_map.get(tu.tool_use_id)
+            if result is None:
+                continue
+            key = _normalize_key(tu.tool_name, tu.tool_input)
+            records.append({
+                "tool_name": tu.tool_name,
+                "key": key,
+                "turn": turn.turn_number,
+                "is_error": _is_error(result),
+            })
+
+    groups: dict[tuple, list] = defaultdict(list)
+    for r in records:
+        groups[(r["tool_name"], r["key"])].append(r)
+
+    found: list[dict] = []
+    seen_pairs: set[tuple] = set()
+
+    for (tool_name, failure_key), group in groups.items():
+        errors = [r for r in group if r["is_error"]]
+        if len(errors) < 2:
+            continue
+
+        first_err_turn = errors[0]["turn"]
+        last_err_turn = errors[-1]["turn"]
+
+        intervening = any(
+            r for r in group
+            if not r["is_error"] and first_err_turn < r["turn"] < last_err_turn
+        )
+        if intervening:
+            continue
+
+        fix = next(
+            (
+                r for r in records
+                if r["tool_name"] == tool_name
+                and not r["is_error"]
+                and r["key"] != failure_key
+                and last_err_turn < r["turn"] <= last_err_turn + FIX_WINDOW
+            ),
+            None,
+        )
+        if fix is None:
+            continue
+
+        pair_key = (tool_name, failure_key, fix["key"])
+        if pair_key in seen_pairs:
+            continue
+        seen_pairs.add(pair_key)
+
+        found.append({
+            "tool_name": tool_name,
+            "failure_key": failure_key,
+            "fix_key": fix["key"],
+            "first_failure_turn": first_err_turn,
+            "fix_turn": fix["turn"],
+        })
+
+    return found
+
+
+def _compute_waste(trace: SessionTrace, first_failure_turn: int, fix_turn: int) -> float:
+    price = price_per_tok(trace.primary_model)
+    total = 0.0
+    for turn in trace.turns:
+        if (
+            turn.role == "assistant"
+            and first_failure_turn <= turn.turn_number <= fix_turn
+            and turn.usage is not None
+        ):
+            total += turn.usage.input_tokens * price
+    return round(total, 4)
+
+
+def detect(pairs: list[tuple[Diagnosis, SessionTrace]]) -> list[ProjectPattern]:
+    """Detect recurring failure/fix patterns across sessions."""
+    session_records: list[dict] = []
+    for _diagnosis, trace in pairs:
+        for p in _find_pairs(trace):
+            session_records.append({
+                "session_id": trace.session_id,
+                "tool_name": p["tool_name"],
+                "failure_key": p["failure_key"],
+                "fix_key": p["fix_key"],
+                "first_failure_turn": p["first_failure_turn"],
+                "fix_turn": p["fix_turn"],
+                "trace": trace,
+            })
+
+    groups: dict[tuple, list] = defaultdict(list)
+    for r in session_records:
+        groups[(r["tool_name"], r["failure_key"], r["fix_key"])].append(r)
+
+    result: list[ProjectPattern] = []
+    for (tool_name, failure_key, fix_key), records in groups.items():
+        seen: dict[str, dict] = {}
+        for r in records:
+            if r["session_id"] not in seen:
+                seen[r["session_id"]] = r
+
+        if len(seen) < MIN_SESSIONS:
+            continue
+
+        unique = list(seen.values())
+        wasted = [r["fix_turn"] - r["first_failure_turn"] for r in unique]
+        avg_wasted_turns = sum(wasted) / len(wasted)
+        total_waste_usd = sum(
+            _compute_waste(r["trace"], r["first_failure_turn"], r["fix_turn"])
+            for r in unique
+        )
+
+        result.append(ProjectPattern(
+            tool_name=tool_name,
+            failure_key=failure_key,
+            fix_key=fix_key,
+            session_count=len(seen),
+            avg_wasted_turns=round(avg_wasted_turns, 1),
+            total_waste_usd=round(total_waste_usd, 4),
+            example_sessions=sorted(r["session_id"] for r in unique)[:3],
+        ))
+
+    return result