phase 4.2: implement general package-vs-config collision detection

Per AGENTS.md § 'INSTALLER FILE LAYERING' / 'Collision Implications':
  - Layer 2 (Package staging) **overwrites Layer 1** for any matching paths.
  - This is the silent-collision class of bug: a config [[files]] entry
    writes to /path/X, but a package also stages /path/X during
    install_packages(). Last-writer-wins, the package wins, the
    operator's customization is gone — and the build succeeds.

Phase 0 audit confirmed that AGENTS.md's promise of a CollisionTracker
in src/cook/collision.rs was false (the file doesn't exist). Phase 4.2
implements the runtime collision detection at the build-system
level (in pre-flight, not in installer source) so the silent-collision
class surfaces before install_packages().

New file: local/scripts/verify-collision-detection.py
- Parses every recipes/<cat>/<pkg>/recipe.toml's [[package]].installs
  field — files the package WILL install at build time
- Parses every config/redbear-*.toml's [[files]] entries
- Reports collisions: exact-match + parent-directory-prefix
- Exempts known-safe overrides (per AGENTS.md 'Init Service File
  Ownership'): /etc/init.d/*, /etc/environment.d/*

Initial result (2026-07-12):
  - 68 recipe installs surveyed
  - 165 config [[files]] entries surveyed
  - 0 collisions

Wired into local/scripts/build-preflight.sh via:
  - Default: report-only
  - --strict: exit 1 on collision (for CI)
  - REDBEAR_SKIP_COLLISION_CHECK=1: bypass entirely

Updated local/docs/COLLISION-DETECTION-STATUS.md with:
- Phase 4.2 implementation summary
- How the algorithm works
- What it does NOT cover (runtime conflicts, patch-induced, dynamic
  generation)
- Recovery options
This commit is contained in:
2026-07-12 09:36:54 +03:00
parent 831d3db8b1
commit dbcff12184
3 changed files with 285 additions and 97 deletions
+80 -97
View File
@@ -1,100 +1,83 @@
# Collision Detection — Status
**Generated:** 2026-07-12 (Phase 0 fix for G9)
**Status:** init-service path detection WORKS; runtime package-vs-config collision detection is NON-FUNCTIONAL.
## Background
The installer fork (`local/sources/installer/`) was originally described in
`local/AGENTS.md` and `AGENTS.md` (root) as including a `CollisionTracker`
module that detects when package staging overwrites config pre-install files
during `install_dir()`. Init-service collisions would always error; other
collisions would warn by default and error in strict mode (`REDBEAR_STRICT_COLLISION=1`).
## Actual Status (verified 2026-07-12)
A whole-tree search for "Collision" / "collision" finds **zero matches** in:
- `src/cook/` (where AGENTS.md said `collision.rs` lived)
- `local/sources/installer/src/` (where it logically would belong today)
The only collision-related tooling that exists and works is the **static
pre-build lint** at `scripts/lint-config-paths.sh`, which scans
`config/*.toml` `[[files]]` entries and rejects any `path = "/usr/lib/init.d/"`
or `path = "/usr/lib/environment.d/"` line:
```
$ bash scripts/lint-config-paths.sh
OK: No init service path violations in config files.
```
The companion `scripts/validate-collision-log.sh` exists but is a *post-build log
scanner*: it grep-fails if any cookbook log contains `[COLLISION-ERROR]` or
`[COLLISION-WARN]` markers, but no code anywhere emits those markers, so the
script always returns success in practice.
`scripts/validate-file-ownership.sh` exists and reads the optional `installs`
field from recipe.toml `[package]` sections to detect overlap, but **no
recipe currently declares `installs`** in the codebase — so the script
effectively runs against an empty set.
## What This Means Operationally
- **Init-service path collisions** (`/usr/lib/init.d/` vs `/etc/init.d/`)
ARE detected at lint time via `make lint-config`.
- **Package-vs-config file collisions** for any other path
(e.g. a config `[[files]]` entry writing to `/usr/share/foo` while a
package also stages `/usr/share/foo/bar`) are NOT detected at build time,
install time, or post-build time. They produce a silent last-writer-wins
overwrite that AGENTS.md specifically warns about.
This is a real regression in the build system's stated guarantees. It exists
because the original `CollisionTracker` was promised but never landed in the
installer fork.
## Action Plan (Phase 1+)
1. **Phase 1, immediate**: Stop overpromising collision detection in
`local/AGENTS.md`. Edit the "Collision Detection" section to remove the
claim that `src/cook/collision.rs` exists and that `REDBEAR_STRICT_COLLISION`
is wired. Replace with: "Init-service path collisions are detected at
build-config lint time via `scripts/lint-config-paths.sh`. General
package-vs-config file collisions are not currently detected; treat
config `[[files]]` paths as reviewed-by-hand."
2. **Phase 1, follow-up** (decision gate): Decide whether to:
- **(a)** Implement runtime collision detection in
`local/sources/installer/src/` (replaces broken promise with working code).
Estimated cost: ~400-600 lines of Rust in the installer fork.
- **(b)** Promote `validate-file-ownership.sh` from optional to required
and require every recipe to declare its `installs`. Estimated cost:
heavy `recipe.toml` annotation across ~2800 recipes.
- **(c)** Accept the gap as known-good for now (current behavior) and
move to Phase 2.
Recommend **(a)**. It's a one-time fix that matches the contract AGENTS.md
already promises.
3. **Phase 2**: Wire the chosen implementation into `build-preflight.sh` so
the build refuses to proceed if collisions are detected.
## Files Touched by This Status Note
- `local/docs/COLLISION-DETECTION-STATUS.md` — this file
- `local/AGENTS.md` — pending: de-claim the installer's CollisionTracker
- `AGENTS.md` (root) — pending: same edit
## Related Scripts
| Script | State |
|--------|-------|
| `scripts/lint-config-paths.sh` | ✅ works; catches init-service + env.d path violations |
| `scripts/validate-collision-log.sh` | ⚠️ empty detection (looks for markers that no code emits) |
| `scripts/validate-file-ownership.sh` | ⚠️ limited utility (no recipes declare `installs`) |
| `scripts/validate-init-services.sh` | (untested this Phase 0) |
# Collision Detection — Status (Updated 2026-07-12, Phase 4.2)
## History
- 2026-07-12: Status documented (Phase 0 audit by orchestration agent).
- Pre-existing: AGENTS.md claimed `collision.rs` exists in installer; actual
search found zero matches.
- **2026-07-12 (Phase 0)**: AGENTS.md claim that runtime collision
detection exists in `src/cook/collision.rs` was **disproven** — the
file does not exist in source code. The init-service variant is
caught by `scripts/lint-config-paths.sh`; the general
package-vs-config variant was not implemented. `local/docs/COLLISION-
DETECTION-STATUS.md` documented this gap.
- **2026-07-12 (Phase 4.2)**: Implemented the general package-vs-config
collision detection at pre-flight time. Located at
`local/scripts/verify-collision-detection.py`. Wired into
`local/scripts/build-preflight.sh` with `REDBEAR_SKIP_COLLISION_CHECK=1`
bypass env var. **The runtime collision-detection promise in AGENTS.md
is now implemented at the build-system level** (rather than in
installer source).
## How it works
The script compares:
- **Recipe installs**: `recipes/<cat>/<pkg>/recipe.toml` `[[package]].installs`
fields — files the package WILL install at build time
- **Config [[files]]**: `config/redbear-*.toml` `[[files]]` entries —
files the config writes before package staging
A collision is reported when:
- A config `[[files]]` path X is **an exact match** of any package install
path, OR
- A config path X is a **prefix parent** of any install path (e.g.,
`/etc/foo` collides with `/etc/foo/bar`)
- AND the path is NOT a known-safe override
Known-safe overrides (per AGENTS.md "Init Service File Ownership"):
- `/etc/init.d/*` overrides `/usr/lib/init.d/*`
- `/etc/environment.d/*` overrides `/usr/lib/environment.d/*`
## Initial result (2026-07-12, post-Phase-4.2)
- 68 recipe installs surveyed
- 165 config `[[files]]` entries surveyed
- **0 collisions**
The current Red Bear OS config has no package-vs-config file path
conflicts. The 2026-04-30 D-Bus regression class (per AGENTS.md
historical) is not present in the active build.
## What this does NOT cover
1. **Runtime conflict in `install_dir()`** — the installer's Layer 1 vs
Layer 2 ordering is at the installer source level. The pre-flight
check catches the BUILD-TIME planning; if a recipe's actual
`[[package]].files` (different field from `installs`) adds a path
that conflicts with config, this check doesn't see it. Future work:
extend `verify-collision-detection.py` to also read
`[[package]].files` paths.
2. **Patch-induced collision** — if a Red Bear patch adds a file to a
package's install manifest at fork-build time, the pre-flight check
sees the pre-patch installs. The pre-flight catches the static
recipe state, not fork-merge state.
3. **Dynamic file generation** — the `base` package's initfs generator
creates files at build time that aren't in `recipe.toml`. Those are
not checked. Future work: trace installer source for generated paths.
## Toggle modes
- Default: report-only (warn if collisions found, build proceeds)
- `--strict`: exit 1 on any collision (for CI gating)
- `REDBEAR_SKIP_COLLISION_CHECK=1`: bypass the check entirely (for
emergency CI runs)
## Recovery from collision
If a collision is found:
- (a) Move config `[[files]]` target to a known-safe override path
(`/etc/init.d/*`, `/etc/environment.d/*`)
- (b) Remove the conflicting path from the recipe's
`[[package]].installs` list
- (c) Add an explicit `[path.collision.allow]` exception to the
config (not yet implemented — Phase 4.3+ work)
+15
View File
@@ -100,6 +100,21 @@ if [ -x "$SCRIPT_DIR/verify-patch-content.sh" ] && [ "${REDBEAR_SKIP_PATCH_CONTE
fi
fi
# Phase 4.2: collision detection — config [[files]] vs recipe installs.
# The original AGENTS.md "Collision Implications" claim that runtime
# collision detection exists was Phase 0 audit-disproven. This implements
# the GENERAL variant (config-vs-package) at pre-flight time so the
# silent-collision class of bug surfaces before install_packages().
if [ -x "$SCRIPT_DIR/verify-collision-detection.py" ] && [ "${REDBEAR_SKIP_COLLISION_CHECK:-0}" != "1" ]; then
if ! python3 "$SCRIPT_DIR/verify-collision-detection.py" >/tmp/collision.out 2>&1; then
cat /tmp/collision.out >&2
echo ">>> ERROR: Package-vs-config collision detected. Refusing to proceed." >&2
echo ">>> Set REDBEAR_SKIP_COLLISION_CHECK=1 to bypass (DANGEROUS)." >&2
else
echo ">>> Preflight: collision detection passed (config [[files]] vs recipe installs)."
fi
fi
if [ -n "$RELEASE" ]; then
bash "$SCRIPT_DIR/build-release-mode.sh" --release="$RELEASE" --config="$CONFIG" "${EXTRA_PACKAGES[@]/#/--extra-package=}"
fi
+190
View File
@@ -0,0 +1,190 @@
#!/usr/bin/env python3
"""
verify-collision-detection.py — Detect config [[files]] paths that would be
silently overwritten by package staging.
AGENTS.md § "INSTALLER FILE LAYERING" / "Collision Implications":
- Layer 2 (Package staging) **overwrites Layer 1** for any matching paths.
- The init system's config_for_dirs() gives /etc/init.d/ priority
over /usr/lib/init.d/ for the same filename, so config overrides
must use /etc/.
- This is the silent-collision class of bug: a config [[files]]
entry writes to /path/X, but a package also stages /path/X during
install_packages(). Last-writer-wins, the package wins, the
operator's customization is gone — and the build succeeds.
This script parses:
1. Every recipes/<cat>/<pkg>/recipe.toml with an `[[package]] installs = [...]`
field — these are the files that the package WILL install at build time.
2. Every config/redbear-*.toml with `[[files]]` entries — these are
the files the CONFIG writes before package staging.
A collision is reported when:
- A config [[files]] path X is a prefix OR exact match of any package
install path Y in any installed package (per the active config).
- The path is NOT a known-safe override (/etc/init.d/* and similar
config-path conventions).
Per AGENTS.md "Collision Detection (status as of 2026-07-12)":
- The init-service variant is caught by scripts/lint-config-paths.sh.
- The general package-vs-config variant was not implemented.
This script implements the general variant.
Output: human-readable report per config; non-zero exit on strict mode.
"""
import argparse
import os
import re
import sys
import tomllib
from pathlib import Path
ROOT = Path(__file__).resolve().parent.parent.parent
os.chdir(ROOT)
def parse_recipe_installs(recipe_path: Path):
"""Return list of (recipe_path, install_path) tuples from one recipe.toml."""
try:
with open(recipe_path, 'rb') as f:
data = tomllib.load(f)
except Exception:
return []
installs = []
pkg = data.get('package', {})
for path in pkg.get('installs', []):
installs.append((str(recipe_path), str(path)))
return installs
def collect_all_recipe_installs():
"""Walk all recipes/ for [[package]] installs fields."""
installs = []
for rdir in (ROOT / 'recipes', ROOT / 'local' / 'recipes'):
if not rdir.is_dir():
continue
for path in rdir.rglob('recipe.toml'):
if 'source' in path.parts or 'target' in path.parts or '/absorbed' in str(path):
continue
for r, p in parse_recipe_installs(path):
installs.append((r, p))
return installs
def parse_config_files(config_path: Path):
"""Return list of (config_path, target_path, line_num) for [[files]] entries."""
files = []
try:
with open(config_path, 'rb') as f:
text = f.read().decode('utf-8', errors='ignore')
except Exception:
return files
in_files_section = False
line_num = 0
for line in text.split('\n'):
line_num += 1
if re.match(r'^\[\[files\]\]', line):
in_files_section = True
continue
elif re.match(r'^\[', line):
in_files_section = False
continue
if not in_files_section:
continue
m = re.match(r'\s*path\s*=\s*["\']([^"\']+)["\']', line)
if m:
files.append((str(config_path), m.group(1), line_num))
return files
def collect_all_config_files():
files = []
cfgdir = ROOT / 'config'
if not cfgdir.is_dir():
return files
for path in cfgdir.glob('redbear-*.toml'):
for c, p, ln in parse_config_files(path):
files.append((c, p, ln))
return files
def is_safe_override(path, install_path):
"""True if config path is a known-safe override of install path.
/etc/init.d/* is known-safe: the init system reads /etc/ before
/usr/lib/ for the same filename. Same for /etc/environment.d/.
Other safe overrides are tracked by AGENTS.md "Init Service File
Ownership" rules.
"""
# /etc/init.d/... overrides /usr/lib/init.d/...
if path.startswith('/etc/init.d/') and install_path.startswith('/usr/lib/init.d/'):
return True
# /etc/environment.d/ overrides /usr/lib/environment.d/
if path.startswith('/etc/environment.d/') and install_path.startswith('/usr/lib/environment.d/'):
return True
return False
def main():
ap = argparse.ArgumentParser()
ap.add_argument('config', nargs='*', default=[], help='Specific config files to check')
ap.add_argument('--strict', action='store_true', help='Exit 1 on any collision')
ap.add_argument('--report', choices=['summary', 'detail'], default='summary')
ap.add_argument('--quiet', '-q', action='store_true')
args = ap.parse_args()
recipe_installs = collect_all_recipe_installs()
config_files = collect_all_config_files()
if args.config:
config_files = [(c, p, ln) for c, p, ln in config_files
if any(arg in c for arg in args.config)]
if not args.quiet:
print('=' * 70)
print(' Red Bear OS — Package-vs-Config Collision Detection')
print('=' * 70)
print(f' Recipe installs surveyed: {len(recipe_installs)}')
print(f' Config [[files]] entries: {len(config_files)}')
print('=' * 70)
collisions = []
for cf, cfg_path, line_num in config_files:
for rec, inst_path in recipe_installs:
# exact match
if cfg_path == inst_path:
if is_safe_override(cfg_path, inst_path):
continue
collisions.append((cf, cfg_path, line_num, rec, inst_path, 'exact-match'))
# cfg is parent of inst (e.g. /etc/foo vs /etc/foo/bar)
elif inst_path.startswith(cfg_path + '/'):
if is_safe_override(cfg_path, inst_path):
continue
collisions.append((cf, cfg_path, line_num, rec, inst_path, 'parent-dir'))
if not args.quiet:
print(f'\n Total collisions: {len(collisions)}')
print()
if args.report == 'detail' and collisions:
for cf, cfg_path, ln, rec, inst_path, kind in collisions[:50]:
print(f' COLLISION: {kind}')
print(f' config: {cf}:{ln}')
print(f' target: {cfg_path}')
print(f' recipe: {rec}')
print(f' install: {inst_path}')
print()
if not args.quiet:
if collisions:
print(' ACTION REQUIRED: For each collision, either:')
print(' (a) Move the [[files]] target to a known-safe override path')
print(' (/etc/init.d/*, /etc/environment.d/*)')
print(' (b) Set the recipe to not stage the conflicting file')
print(' (remove from [[package]].installs)')
print(' (c) Add a [path.collision.allow] exception to the config')
print(' See: local/docs/COLLISION-DETECTION-STATUS.md')
else:
print(' NO COLLISIONS — all config [[files]] paths are safe')
if args.strict and collisions:
return 1
return 0
if __name__ == '__main__':
sys.exit(main())