codegraph-brain 0.7.1__tar.gz → 0.7.2__tar.gz

{codegraph_brain-0.7.1 → codegraph_brain-0.7.2}/.github/workflows/guardian.yml

@@ -67,7 +67,7 @@ jobs:
         uses: astral-sh/setup-uv@caf0cab7a618c569241d31dcd442f54681755d39  # v3
 
       - name: Install dependencies
-        run: uv sync --group guardian
+        run: uv sync --extra guardian
 
       - name: Build code graph
         run: uv run cgis ingest ./src --output graph.db

codegraph_brain-0.7.2/.release-please-manifest.json

@@ -0,0 +1,3 @@
+{
+  ".": "0.7.2"
+}

{codegraph_brain-0.7.1 → codegraph_brain-0.7.2}/CHANGELOG.md

@@ -1,5 +1,19 @@
 # Changelog
 
+## [0.7.2](https://github.com/zaebee/codegraph-brain/compare/codegraph-brain-v0.7.1...codegraph-brain-v0.7.2) (2026-07-31)
+
+
+### Bug Fixes
+
+* **ci:** install guardian deps with --extra, not --group ([#300](https://github.com/zaebee/codegraph-brain/issues/300)) ([1da003d](https://github.com/zaebee/codegraph-brain/commit/1da003d811db8e60201693905a6f83c2fccf48cf))
+* **cli:** escape data interpolated into Rich console output ([#298](https://github.com/zaebee/codegraph-brain/issues/298)) ([dd4943d](https://github.com/zaebee/codegraph-brain/commit/dd4943d0fea1100b47130fcb0790e940da83276a))
+* **drift:** pin measured depth when a domain is shallower than its gate ([#297](https://github.com/zaebee/codegraph-brain/issues/297)) ([1ebb93a](https://github.com/zaebee/codegraph-brain/commit/1ebb93a888037af5ad333ddffc7896af61f6eb5e)), closes [#229](https://github.com/zaebee/codegraph-brain/issues/229)
+
+
+### Documentation
+
+* add the chunked-review lab note (a negative result worth keeping) ([#299](https://github.com/zaebee/codegraph-brain/issues/299)) ([265fefd](https://github.com/zaebee/codegraph-brain/commit/265fefda7ee30ece692d13ea0bb171d99ac18bd9)), closes [#160](https://github.com/zaebee/codegraph-brain/issues/160)
+
 ## [0.7.1](https://github.com/zaebee/codegraph-brain/compare/codegraph-brain-v0.7.0...codegraph-brain-v0.7.1) (2026-07-31)
 
 

{codegraph_brain-0.7.1 → codegraph_brain-0.7.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: codegraph-brain
-Version: 0.7.1
+Version: 0.7.2
 Summary: Semantic code graph for AI agents — deterministic FQN resolution, impact analysis and architectural drift gates, exposed over MCP.
 Project-URL: Homepage, https://github.com/zaebee/codegraph-brain
 Project-URL: Repository, https://github.com/zaebee/codegraph-brain

codegraph_brain-0.7.2/docs/lab-notes/2026-06-11-chunked-review-negative-result.md

@@ -0,0 +1,132 @@
+# Lab note: the chunked-review experiment — a negative result worth keeping
+
+**Date:** 2026-06-11
+**Experiment:** #154 slices 1–2 · **Code:** PR #157 (`6178bae`), PR #159 (`2030eb8`)
+**Verdict:** gate FAILED 3/4 — feature stays behind a flag, off in production
+**Specs:** [`2026-06-11-guardian-chunker-design.md`](../specs/2026-06-11-guardian-chunker-design.md) ·
+[`2026-06-11-guardian-chunked-review-design.md`](../specs/2026-06-11-guardian-chunked-review-design.md)
+
+The headline: the hypothesis survived contact with reality only partially, and the
+honest negative result is the most valuable artifact.
+
+## The problem
+
+Guardian (our in-repo LLM reviewer, replacing gemini-code-assist before its
+2026-07-17 sunset) reliably LGTMs large PRs. Baseline bench (6 replayed PRs with
+hand-curated ground truth, 3 runs each): mean recall 0.22, but **recall = 0 on all
+three largest PRs** (36–62K prompt tokens). Plan-2 ablations proved more context
+doesn't help: feeding 94K tokens of full files still produced LGTMs.
+
+Diagnosis: **attention dilution** — the finder skims, it doesn't read.
+
+## The hypothesis (#154)
+
+If the finder can't read a big diff, give it small ones: split changed files into
+connected components over the code graph (IMPORTS/CALLS from `cgis ingest`), run a
+finder pass per chunk with a per-chunk context budget, merge, dedup, run ONE
+skeptic pass, post one report. Each chunk = a small, complete world.
+
+## What we built
+
+- **Slice 1 (PR #157, `6178bae`)** — `chunker.py`, pure logic. Per-file diff
+  splitting (rename/deletion/binary/git-quoted-path handling), union-find over
+  graph edges, test-pairing heuristic, deterministic ordering, honest degradation
+  to isolated chunks when the graph DB is missing.
+- **Slice 2 (PR #159, `2030eb8`)** — wiring behind `GUARDIAN_FEATURES=chunked`.
+  `run_review_routed` entry point, per-chunk finder passes with per-chunk
+  `full_files` (120K char budget) + impact graph, out-of-chunk hallucination
+  filter with LLM-path-artifact normalization (`a/x.py` → `x.py`),
+  `(file,line,category)` dedup, single skeptic over finding-bearing chunks only,
+  `MAX_CHUNKS=8` cap with overflow merge, `cumulative_usage` token accounting
+  (which incidentally fixed a latent retry-loses-tokens bug).
+
+Both slices SDD-executed (subagent implementer + spec reviewer + quality
+reviewer), bench-gated by spec §7 **before** enabling anything in prod.
+
+## The gate (pre-registered, spec §7)
+
+1. mean recall (6 PRs) ≥ 0.27 — baseline + 0.05
+2. mean recall over large PRs {140, 143, 144} > 0 — the hypothesis must show where it matters
+3. mean noise ≤ 1.5/PR
+4. no PR drops > 0.05 below its baseline
+
+## The result: FAIL 3/4
+
+18 runs (6 PRs × 3, finder `gemini-2.5-flash`, skeptic `gemini-3.5-flash`), data in
+`benchmarks/guardian/results.jsonl`:
+
+| PR | chunks | recall base → chunked | noise base → chunked |
+|---|---|---|---|
+| 122 | 8 (cap) | 0.152 → 0.152 | 1.0 → **6.0** |
+| 140 | 5 | 0 → **0.067** ✨ | 3.3 → 3.3 |
+| 141 (noise probe) | 5 | 1.0 → 1.0 | 0.7 → 2.7 |
+| 142 | 1 (degenerate) | 0.167 → 0 | 0 → 0 |
+| 143 | 7 | 0 → 0 | 0 → 2.7 |
+| 144 | 5 | 0 → 0 | 0.7 → 1.3 |
+
+Mean recall 0.203 (❌ vs 0.27), large-PR recall 0.022 (✅ — but only PR 140
+contributed), noise 2.67/PR (❌ vs 1.5), PR 142 regression (❌, with a variance
+caveat: its baseline rested on one lucky 0.5 run out of three).
+
+## What we learned
+
+1. **Attention dilution is real but not the whole story.** Chunking produced the
+   first-ever nonzero recall on a large PR (140). But 143/144 stayed at zero across
+   21 small, focused finder calls. Those ground-truth bugs are invisible to this
+   finder *regardless of context size* — a model/prompt capability ceiling, not a
+   context problem.
+2. **Chunking multiplies noise superlinearly.** Every chunk invents its own false
+   positives (122: 8 chunks → noise 6.0/run). A single skeptic pass over
+   concatenated chunk contexts can't compensate.
+3. **Pre-registered gates work.** Writing pass/fail criteria into the spec before
+   spending money made the no-go decision mechanical instead of motivated. The flag
+   stays OFF in prod; the code sleeps behind it, fully tested, waiting for a better
+   noise story.
+4. **Honest degradation pays off.** Every failure mode (no graph DB, flaky chunk,
+   unparsable finder output, dead skeptic) degrades to something useful instead of
+   crashing — which also made the bench itself survivable.
+
+## Where this goes next (slice 3 candidates, not committed)
+
+- Noise mitigation: stricter per-chunk finder prompt, per-chunk findings cap,
+  skeptic hardening for chunked volume.
+- Ensemble finder (#155): union of N runs matched *different* GT findings
+  (122: 0.09/0.09/0.18 → union ~0.27) — possibly a cheaper recall lever than chunking.
+- Eyeball the 143/144 ground truths: if they need cross-file reasoning the finder
+  can't do, no amount of context plumbing will fix it.
+- Quotient-level "architectural chunk" (domain-graph drift as its own review unit).
+
+Total experiment cost: ~5 PLN of gemini billing for the slice-2 bench (1.03M prompt
+tokens), on top of Plan-2's bench budget. Money well burned.
+
+## Reproducing these numbers
+
+`benchmarks/guardian/results.jsonl` is committed, so the chunked side of this
+experiment is re-derivable. Group rows by `pr`, take the 3 rows whose `features`
+field contains `chunked` (all share `guardian_sha` `a866a2ac`), and average per PR:
+
+| Figure | Note | Recomputed |
+|---|---|---|
+| mean recall (6 PRs) | 0.203 | 0.203 |
+| mean noise / PR | 2.67 | 2.667 |
+| large-PR recall {140,143,144} | 0.022 | 0.022 |
+
+Every cell of the table's **chunked** column reproduces exactly.
+
+**The `base` column does not.** No single `guardian_sha` slice of the
+`features == ""` rows reproduces it, and the baseline headline of 0.22 mean recall
+does not fall out of any of them either — the two candidate slices with exactly 3
+runs per PR give 0.191 and 0.167, and the earliest slice gives 0.271 on uneven run
+counts. The file has also been appended to since (there are July 2026 rows from
+later work), so the working set used to compute the baseline in June is no longer
+recoverable from it.
+
+This is recorded rather than corrected: a lab journal is a historical record, and
+silently restating the baseline would be worse than flagging it. The gate
+conclusions do not rest on it — gates 1 and 3 failed on absolute values (0.203 <
+0.27, 2.67 > 1.5), and both reproduce. Gate 4 (per-PR regression vs baseline) is
+the one that cannot be independently re-checked today.
+
+## Refs
+
+#154, #155, #160 · PR #157, PR #159

codegraph_brain-0.7.2/docs/ontology/tolerances.lock

@@ -0,0 +1,27 @@
+# Drift-tolerance ratchet lock (#151) — generated from patterns.yaml, edited by hand.
+#
+# Project convention since #140/#141: drift_tolerance is a RATCHET. Values may
+# only go DOWN as domains get healthier, never up to mask a regression.
+#
+# test_tolerances_never_exceed_the_lock enforces it: lowering a tolerance means
+# updating this file in the same PR, which puts the change in the diff where a
+# reviewer sees it. Raising one fails the suite.
+#
+# Deliberate re-baselining (e.g. the Part B re-baseline in #144) is allowed —
+# update the lock and justify it in the PR description.
+
+project_domains:
+  extraction: 0.40
+  resolution: 0.40
+  pipeline: 0.25
+  storage: 0.20
+  query: 0.20
+  components: 0.20
+  layout: 0.50
+  hooks: 0.15
+  store: 0.15
+  providers: 0.15
+  utils: 0.15
+
+project_level:
+  cgis-project: 0.15

{codegraph_brain-0.7.1 → codegraph_brain-0.7.2}/pyproject.toml

@@ -7,7 +7,7 @@ packages = ["src/cgis"]
 
 [project]
 name = "codegraph-brain"
-version = "0.7.1"
+version = "0.7.2"
 description = "Semantic code graph for AI agents — deterministic FQN resolution, impact analysis and architectural drift gates, exposed over MCP."
 readme = "README.md"
 requires-python = ">=3.12"
@@ -81,6 +81,7 @@ dev = [
     "interrogate>=1.7.0",
     "ipdb>=0.13.13",
     "mypy>=2.1.0",
+    "networkx>=3.3",
     "pre-commit>=4.6.0",
     "pytest>=9.0.3",
     "pytest-asyncio>=0.24.0",

{codegraph_brain-0.7.1 → codegraph_brain-0.7.2}/src/cgis/cli.py

@@ -1,4 +1,22 @@
-""" "CLI to run pipeline."""
+""" "CLI to run pipeline.
+
+Rich markup convention (#148)
+-----------------------------
+Rich parses ``[...]`` in everything it renders, so any *data* reaching a console
+sink must be wrapped in ``rich.markup.escape``.  This applies to ``console.print``
+f-strings, ``Table.add_row`` cells, and panel/tree labels alike.
+
+The failure mode is silent: unescaped text is not styled, it *vanishes*.  PR #144
+shipped a line where ``[cgis-project]`` rendered as an empty string, so the
+operator saw a blank where their domain name should have been.
+
+  wrong:  console.print(f"[bold red]Not found:[/bold red] {path}")
+  right:  console.print(f"[bold red]Not found:[/bold red] {escape(path)}")
+
+Escape anything derived from config, the graph, the filesystem, CLI arguments, or
+an exception message — names, FQNs, file paths, pattern names, ``{e}``.  Literal
+style tags you wrote yourself and pure numbers need no escaping.
+"""
 
 import dataclasses
 import json as _json
@@ -190,10 +208,12 @@ def ingest(
     pipeline = IngestionPipeline(extractors, domains_config=domains)
 
     if domains and not Path(domains).is_file():
-        console.print(f"[bold red]❌ Domains config file not found:[/bold red] {domains}")
+        console.print(
+            f"[bold red]❌ Domains config file not found:[/bold red] {escape(str(domains))}"
+        )
         raise typer.Exit(code=1)
 
-    console.print(f"[bold blue]🚀 Starting ingestion for:[/bold blue] {path}")
+    console.print(f"[bold blue]🚀 Starting ingestion for:[/bold blue] {escape(path)}")
 
     if incremental and output.endswith(".json"):
         console.print(
@@ -233,7 +253,7 @@ def ingest(
         console.print("[bold green]✅ Success![/bold green] Graph data ready.")
 
     except Exception as e:
-        console.print(f"[bold red]❌ Error during ingestion:[/bold red] {e}")
+        console.print(f"[bold red]❌ Error during ingestion:[/bold red] {escape(str(e))}")
         raise typer.Exit(code=1) from e
 
 
@@ -272,16 +292,16 @@ def _resolve_cli_fqn(store: SQLiteStore, target: str, kind: str) -> str:
     resolution = resolve_fqn(store, target)
     if resolution.resolved is None:
         if resolution.candidates:
-            console.print(f"[bold red]❌ Ambiguous FQN:[/bold red] {target}")
+            console.print(f"[bold red]❌ Ambiguous FQN:[/bold red] {escape(target)}")
             for candidate in resolution.candidates:
-                console.print(f"  [dim]- {candidate}[/dim]")
+                console.print(f"  [dim]- {escape(candidate)}[/dim]")
             if resolution.truncated:
                 console.print(_TRUNCATED_HINT)
         else:
-            console.print(f"[bold red]❌ {kind} not found in graph:[/bold red] {target}")
+            console.print(f"[bold red]❌ {kind} not found in graph:[/bold red] {escape(target)}")
         raise typer.Exit(code=1)
     if resolution.via_suffix:
-        console.print(f"[dim]Resolved '{target}' → '{resolution.resolved}'[/dim]")
+        console.print(f"[dim]Resolved '{escape(target)}' → '{escape(resolution.resolved)}'[/dim]")
     return resolution.resolved
 
 
@@ -368,7 +388,9 @@ def trace(
     """
     path = Path(db)
     if not path.is_file():
-        console.print(f"[bold red]❌ Database not found:[/bold red] {db}. Run `ingest` first.")
+        console.print(
+            f"[bold red]❌ Database not found:[/bold red] {escape(db)}. Run `ingest` first."
+        )
         raise typer.Exit(code=1)
 
     allowed: frozenset[EdgeType] | None = None if show_structure else BEHAVIORAL_EDGE_TYPES
@@ -497,7 +519,9 @@ def impact(
     """
     path = Path(db)
     if not path.is_file():
-        console.print(f"[bold red]❌ Database not found:[/bold red] {db}. Run `ingest` first.")
+        console.print(
+            f"[bold red]❌ Database not found:[/bold red] {escape(db)}. Run `ingest` first."
+        )
         raise typer.Exit(code=1)
 
     allowed: frozenset[EdgeType] | None = None if show_structure else BEHAVIORAL_EDGE_TYPES
@@ -562,14 +586,16 @@ def validate(
     """
     path = Path(db)
     if not path.is_file():
-        console.print(f"[bold red]❌ Database not found:[/bold red] {db}. Run `ingest` first.")
+        console.print(
+            f"[bold red]❌ Database not found:[/bold red] {escape(db)}. Run `ingest` first."
+        )
         raise typer.Exit(code=1)
 
     try:
         with SQLiteStore(db) as store:
             stats = store.get_edge_stats()
     except Exception as e:
-        console.print(f"[bold red]❌ Error reading database:[/bold red] {e}")
+        console.print(f"[bold red]❌ Error reading database:[/bold red] {escape(str(e))}")
         raise typer.Exit(code=1) from e
 
     def _pct(n: int) -> str:
@@ -600,7 +626,7 @@ def validate(
         top_table.add_column("count", justify="right", style="yellow")
         for target, count in stats.top_unresolved:
             name = target.removeprefix(RAW_CALL_PREFIX)
-            top_table.add_row(name, str(count))
+            top_table.add_row(escape(name), str(count))
         console.print(top_table)
 
     threshold_pct = threshold * 100
@@ -637,7 +663,9 @@ def find(
     """
     path = Path(db)
     if not path.is_file():
-        console.print(f"[bold red]❌ Database not found:[/bold red] {db}. Run `ingest` first.")
+        console.print(
+            f"[bold red]❌ Database not found:[/bold red] {escape(db)}. Run `ingest` first."
+        )
         raise typer.Exit(code=1)
 
     kinds = (kind.strip().upper(),) if kind and kind.strip() else ()
@@ -654,7 +682,12 @@ def find(
     table.add_column("FQN", style="yellow")
     table.add_column("Location", style="dim")
     for node in matches:
-        table.add_row(node.name, node.type.value, node.id, f"{node.file_path}:{node.start_line}")
+        table.add_row(
+            escape(node.name),
+            node.type.value,
+            escape(node.id),
+            escape(f"{node.file_path}:{node.start_line}"),
+        )
     console.print(table)
 
 
@@ -675,14 +708,16 @@ def structure(
     """
     path = Path(db)
     if not path.is_file():
-        console.print(f"[bold red]❌ Database not found:[/bold red] {db}. Run `ingest` first.")
+        console.print(
+            f"[bold red]❌ Database not found:[/bold red] {escape(db)}. Run `ingest` first."
+        )
         raise typer.Exit(code=1)
 
     # Normalize file path → FQN
     if "/" in target or "\\" in target or target.endswith(".py"):
         normalized = target.replace("\\", "/").removeprefix("./")
         target = file_path_to_module_fqn(normalized)
-        console.print(f"[dim]→ FQN: {target}[/dim]")
+        console.print(f"[dim]→ FQN: {escape(target)}[/dim]")
 
     with SQLiteStore(db) as store:
         target = _resolve_cli_fqn(store, target, "Node")
@@ -696,7 +731,7 @@ def structure(
         typer.echo(_render_graph(output_format, target, nodes, edges))
         return
 
-    console.print(f"[bold blue]📦 Structure of:[/bold blue] {target}\n")
+    console.print(f"[bold blue]📦 Structure of:[/bold blue] {escape(target)}\n")
     root_label = (
         f"[bold cyan]{target_node.type.value}[/bold cyan] [yellow]{target_node.id}[/yellow]"
     )
@@ -765,7 +800,9 @@ def analyze(
     """
     path = Path(db)
     if not path.is_file():
-        console.print(f"[bold red]❌ Database not found:[/bold red] {db}. Run `ingest` first.")
+        console.print(
+            f"[bold red]❌ Database not found:[/bold red] {escape(db)}. Run `ingest` first."
+        )
         raise typer.Exit(code=1)
 
     with SQLiteStore(db) as store:
@@ -809,8 +846,8 @@ def analyze(
                 f"severity=[bold {colour}]{a.severity_score:.2f}[/bold {colour}]"
             )
             for key, val in a.metrics.items():
-                console.print(f"    [dim]{key}:[/dim] {val}")
-            console.print(f"  [italic]💡 {a.refactoring_hint}[/italic]\n")
+                console.print(f"    [dim]{escape(str(key))}:[/dim] {escape(str(val))}")
+            console.print(f"  [italic]💡 {escape(a.refactoring_hint)}[/italic]\n")
 
 
 _DEFAULT_METRICS = "guardian_metrics.jsonl"
@@ -827,7 +864,7 @@ def guardian_rate(
     if updated:
         console.print(f"[green]✅ PR #{pr}: recorded {applied} applied findings.[/green]")
     else:
-        console.print(f"[red]❌ No unrated entry found for PR #{pr} in {metrics}.[/red]")
+        console.print(f"[red]❌ No unrated entry found for PR #{pr} in {escape(metrics)}.[/red]")
         raise typer.Exit(code=1)
 
 
@@ -839,7 +876,7 @@ def guardian_stats(
     """Show Guardian review quality metrics trend."""
     reviews = load_reviews(Path(metrics))
     if not reviews:
-        console.print(f"[yellow]No metrics found in {metrics}.[/yellow]")
+        console.print(f"[yellow]No metrics found in {escape(metrics)}.[/yellow]")
         raise typer.Exit
 
     reviews = reviews[-last:]
@@ -871,7 +908,7 @@ def guardian_stats(
         lgtm = "✅" if r.get("lgtm") else ""
         table.add_row(
             pr_str,
-            str(r.get("model", "")),
+            escape(str(r.get("model", ""))),
             f"{tokens:,}",
             str(findings),
             applied_str,
@@ -962,8 +999,8 @@ def _render_drift_table(reports: list[DriftReport]) -> None:
     table.add_column("Status", justify="center")
     for r in reports:
         table.add_row(
-            r.fqn_prefix,
-            r.expected_pattern or "(hygiene)",
+            escape(r.fqn_prefix),
+            escape(r.expected_pattern or "(hygiene)"),
             f"{r.drift_score:.2f}",
             f"{r.tv_imports:.2f}" if r.tv_imports is not None else "—",
             f"{r.tv_calls:.2f}" if r.tv_calls is not None else "—",
@@ -1029,11 +1066,13 @@ def drift(
     Exits with code 1 if any domain drift score meets or exceeds the critical threshold.
     """
     if not Path(db).is_file():
-        console.print(f"[bold red]❌ Database not found:[/bold red] {db}. Run `ingest` first.")
+        console.print(
+            f"[bold red]❌ Database not found:[/bold red] {escape(db)}. Run `ingest` first."
+        )
         raise typer.Exit(code=1)
 
     if not Path(patterns).is_file():
-        console.print(f"[bold red]❌ Patterns file not found:[/bold red] {patterns}")
+        console.print(f"[bold red]❌ Patterns file not found:[/bold red] {escape(patterns)}")
         raise typer.Exit(code=1)
 
     try:
@@ -1041,7 +1080,7 @@ def drift(
             db, patterns, max_drift=max_drift, profile=profile, max_residual=max_residual
         )
     except Exception as e:
-        console.print(f"[bold red]❌ Error during drift analysis:[/bold red] {e}")
+        console.print(f"[bold red]❌ Error during drift analysis:[/bold red] {escape(str(e))}")
         raise typer.Exit(code=1) from e
 
     if output_format == DriftOutputFormat.JSON:
@@ -1123,17 +1162,19 @@ def init_ontology(
 ) -> None:
     """Propose a starter patterns.yaml from the measured graph (measure-then-label)."""
     if Path(out).exists() and not force:
-        console.print(f"[bold red]❌ {out} already exists[/bold red] — use --force to overwrite.")
+        console.print(
+            f"[bold red]❌ {escape(out)} already exists[/bold red] — use --force to overwrite."
+        )
         raise typer.Exit(code=1)
     try:
         text = propose_ontology(db, margin=margin, min_nodes=min_nodes, depth=depth)
     except FileNotFoundError as e:
-        console.print(f"[bold red]❌ {e}[/bold red]")
+        console.print(f"[bold red]❌ {escape(str(e))}[/bold red]")
         raise typer.Exit(code=1) from e
     Path(out).write_text(text)
     _render_init_summary(text)
-    console.print(f"[bold green]✅ Proposed ontology written to {out}[/bold green]")
-    console.print(f"Next: [cyan]cgis drift --db {db} --patterns {out}[/cyan]")
+    console.print(f"[bold green]✅ Proposed ontology written to {escape(out)}[/bold green]")
+    console.print(f"Next: [cyan]cgis drift --db {escape(db)} --patterns {escape(out)}[/cyan]")
 
 
 @app.command()
@@ -1163,7 +1204,9 @@ def context(
     """
     path = Path(db)
     if not path.is_file():
-        console.print(f"[bold red]❌ Database not found:[/bold red] {db}. Run `ingest` first.")
+        console.print(
+            f"[bold red]❌ Database not found:[/bold red] {escape(db)}. Run `ingest` first."
+        )
         raise typer.Exit(code=1)
 
     err_console = Console(stderr=True)
@@ -1171,16 +1214,18 @@ def context(
         resolution = resolve_fqn(store, fqn)
         if resolution.resolved is None:
             if resolution.candidates:
-                err_console.print(f"[bold red]❌ Ambiguous FQN:[/bold red] {fqn}")
+                err_console.print(f"[bold red]❌ Ambiguous FQN:[/bold red] {escape(fqn)}")
                 for candidate in resolution.candidates:
-                    err_console.print(f"  [dim]- {candidate}[/dim]")
+                    err_console.print(f"  [dim]- {escape(candidate)}[/dim]")
                 if resolution.truncated:
                     err_console.print(_TRUNCATED_HINT)
             else:
-                err_console.print(f"[bold red]❌ Node not found in graph:[/bold red] {fqn}")
+                err_console.print(f"[bold red]❌ Node not found in graph:[/bold red] {escape(fqn)}")
             raise typer.Exit(code=1)
         if resolution.via_suffix:
-            err_console.print(f"[dim]Resolved '{fqn}' → '{resolution.resolved}'[/dim]")
+            err_console.print(
+                f"[dim]Resolved '{escape(fqn)}' → '{escape(resolution.resolved)}'[/dim]"
+            )
         payload = build_context(store, resolution.resolved, depth=depth, source_root=source_root)
     typer.echo(payload)
 
@@ -1193,14 +1238,16 @@ def _render_metrics(report: ArchitectureReport) -> None:
     bottlenecks.add_column("In", justify="right", style="green")
     bottlenecks.add_column("Out", justify="right", style="yellow")
     for m in report.bottlenecks:
-        bottlenecks.add_row(m.node_id, m.node_type, str(m.in_degree), str(m.out_degree))
+        bottlenecks.add_row(
+            escape(m.node_id), escape(m.node_type), str(m.in_degree), str(m.out_degree)
+        )
     console.print(bottlenecks)
 
     gods = Table(title="🏛️  God classes (top by declared members)")
     gods.add_column("Class", style="cyan")
     gods.add_column("Members", justify="right", style="red")
     for m in report.god_classes:
-        gods.add_row(m.node_id, str(m.out_degree))
+        gods.add_row(escape(m.node_id), str(m.out_degree))
     console.print(gods)
 
     critical = Table(title="⭐ Critical nodes (top by PageRank — transitive importance)")
@@ -1213,7 +1260,11 @@ def _render_metrics(report: ArchitectureReport) -> None:
         # In/Out are over the same internal graph PageRank ran on; a high rank with
         # In=0 is a dangling-mass leaf artifact, not a real hub (#237).
         critical.add_row(
-            m.node_id, m.node_type, f"{m.page_rank:.4f}", str(m.in_degree), str(m.out_degree)
+            escape(m.node_id),
+            escape(m.node_type),
+            f"{m.page_rank:.4f}",
+            str(m.in_degree),
+            str(m.out_degree),
         )
     console.print(critical)
 
@@ -1245,7 +1296,9 @@ def metrics(
         console.print("[bold red]❌ metrics supports --format text or json only.[/bold red]")
         raise typer.Exit(code=2)
     if not Path(db).is_file():
-        console.print(f"[bold red]❌ Database not found:[/bold red] {db}. Run `ingest` first.")
+        console.print(
+            f"[bold red]❌ Database not found:[/bold red] {escape(db)}. Run `ingest` first."
+        )
         raise typer.Exit(code=1)
     try:
         with DuckDBAnalyzer(db) as analyzer:
@@ -1253,7 +1306,7 @@ def metrics(
                 bottleneck_limit=limit, god_limit=limit, critical_limit=limit, exclude=exclude
             )
     except Exception as e:  # duckdb missing, extension fetch, or a non-SQLite file
-        console.print(f"[bold red]❌ {e}[/bold red]")
+        console.print(f"[bold red]❌ {escape(str(e))}[/bold red]")
         raise typer.Exit(code=1) from e
 
     if output_format == OutputFormat.JSON:
@@ -1271,16 +1324,20 @@ def _resolve_checkpoint(store: SQLiteStore, target: str) -> str:
     resolution = resolve_fqn(store, target)
     if resolution.resolved is None:
         if resolution.candidates:
-            err_console.print(f"[bold red]❌ Ambiguous checkpoint FQN:[/bold red] {target}")
+            err_console.print(f"[bold red]❌ Ambiguous checkpoint FQN:[/bold red] {escape(target)}")
             for candidate in resolution.candidates:
-                err_console.print(f"  [dim]- {candidate}[/dim]")
+                err_console.print(f"  [dim]- {escape(candidate)}[/dim]")
             if resolution.truncated:
                 err_console.print(_TRUNCATED_HINT)
         else:
-            err_console.print(f"[bold red]❌ Checkpoint not found in graph:[/bold red] {target}")
+            err_console.print(
+                f"[bold red]❌ Checkpoint not found in graph:[/bold red] {escape(target)}"
+            )
         raise typer.Exit(code=1)
     if resolution.via_suffix:
-        err_console.print(f"[dim]Resolved '{target}' → '{resolution.resolved}'[/dim]")
+        err_console.print(
+            f"[dim]Resolved '{escape(target)}' → '{escape(resolution.resolved)}'[/dim]"
+        )
     return resolution.resolved
 
 
@@ -1347,7 +1404,9 @@ def audit(
         )
         raise typer.Exit(code=2)
     if not Path(db).is_file():
-        console.print(f"[bold red]❌ Database not found:[/bold red] {db}. Run `ingest` first.")
+        console.print(
+            f"[bold red]❌ Database not found:[/bold red] {escape(db)}. Run `ingest` first."
+        )
         raise typer.Exit(code=1)
 
     with SQLiteStore(db) as store:
@@ -1434,12 +1493,14 @@ def suggest_packages_cmd(
     always exits 0 on success. Run `ingest` first.
     """
     if not Path(db).is_file():
-        console.print(f"[bold red]❌ Database not found:[/bold red] {db}. Run `ingest` first.")
+        console.print(
+            f"[bold red]❌ Database not found:[/bold red] {escape(db)}. Run `ingest` first."
+        )
         raise typer.Exit(code=1)
     try:
         report = suggest_packages(db, prefix, with_calls=with_calls, min_q=min_q)
     except Exception as e:
-        console.print(f"[bold red]❌ Error during suggest-packages:[/bold red] {e}")
+        console.print(f"[bold red]❌ Error during suggest-packages:[/bold red] {escape(str(e))}")
         raise typer.Exit(code=1) from e
 
     if output_format == SuggestOutputFormat.JSON:
@@ -1472,7 +1533,7 @@ def _render_fractal(reports: list[FractalReport]) -> None:
             entropy = "—" if rung.entropy is None else f"{rung.entropy:.2f}"
             name = rung.name if rung.live else f"{rung.name} [dim](no_signal)[/dim]"
             table.add_row(
-                name,
+                escape(name),
                 str(rung.groups),
                 str(rung.triads),
                 entropy,
@@ -1500,10 +1561,12 @@ def fractal(
     try:
         reports = analyze_fractal_db(db)
     except FileNotFoundError as e:
-        console.print(f"[bold red]❌ Database not found:[/bold red] {db}. Run `ingest` first.")
+        console.print(
+            f"[bold red]❌ Database not found:[/bold red] {escape(db)}. Run `ingest` first."
+        )
         raise typer.Exit(code=1) from e
     except Exception as e:
-        console.print(f"[bold red]❌ Error during fractal analysis:[/bold red] {e}")
+        console.print(f"[bold red]❌ Error during fractal analysis:[/bold red] {escape(str(e))}")
         raise typer.Exit(code=1) from e
 
     if output_format == FractalOutputFormat.JSON: