feat(automation): add bounded overnight lint loop

.gitignore

@@ -7,6 +7,7 @@ coverage/
 .env.local
 tests/e2e/.env.test
 .vitest-cache/
+.codex/night-loop/
 # Local Claude artifacts stay ignored, but shared agent skills are tracked.
 .claude/*
 !.claude/skills/

context/map/artifact-1-territory.md

@@ -2,7 +2,7 @@
 
 <!-- GENERATED by tools/repo-map/scan.mjs — do not hand-edit. Run `npm run repo-map`. -->
 
-> Scan of `7e3dab4` (2026-07-28) · window: last 12 months · engine: tools/repo-map
+> Scan of `c823d6d` (2026-07-28) · window: last 12 months · engine: tools/repo-map
 
 **Where the work actually happens.** Git-history activity over the last 12 months, noise filtered (lockfiles, generated files, assets). Activity is a signal of *risk of accidental regression*, not of quality.
 
@@ -11,11 +11,11 @@
 | Area | Commits |
 | --- | --- |
 | `src/lib` | 75 |
+| `context/map` | 54 |
 | `context/changes` | 53 |
-| `context/map` | 50 |
 | `tests/e2e` | 32 |
 | `src/commands` | 30 |
-| `package.json` | 28 |
+| `package.json` | 29 |
 | `.github/workflows` | 19 |
 | `context/archive` | 19 |
 | `tests/get-command.test.ts` | 13 |
@@ -30,15 +30,15 @@
 
 | File | Commits |
 | --- | --- |
-| `package.json` | 28 |
+| `package.json` | 29 |
 | `src/commands/get.ts` | 16 |
 | `.github/workflows/ci.yml` | 14 |
 | `tests/get-command.test.ts` | 13 |
 | `src/lib/writer.ts` | 12 |
-| `context/map/artifact-1-territory.md` | 10 |
-| `context/map/artifact-2-structure.md` | 10 |
-| `context/map/artifact-3-contributors.md` | 10 |
-| `context/map/repo-map.md` | 10 |
+| `context/map/artifact-1-territory.md` | 11 |
+| `context/map/artifact-2-structure.md` | 11 |
+| `context/map/artifact-3-contributors.md` | 11 |
+| `context/map/repo-map.md` | 11 |
 | `README.md` | 9 |
 | `tests/writer.test.ts` | 9 |
 | `context/map/graph/coupling.json` | 8 |
@@ -51,8 +51,8 @@
 | Area | 2026-Q2 | 2026-Q3 |
 | --- | --- | --- |
 | `src/lib` | 34 | 1 |
+| `context/map` | 0 | 11 |
 | `context/changes` | 28 | 1 |
-| `context/map` | 0 | 10 |
 | `tests/e2e` | 11 | 0 |
 | `src/commands` | 23 | 0 |
 

context/map/artifact-2-structure.md

@@ -2,7 +2,7 @@
 
 <!-- GENERATED by tools/repo-map/scan.mjs — do not hand-edit. Run `npm run repo-map`. -->
 
-> Scan of `7e3dab4` (2026-07-28) · window: last 12 months · engine: tools/repo-map
+> Scan of `c823d6d` (2026-07-28) · window: last 12 months · engine: tools/repo-map
 
 **How it is built.** Static import graph of `src` plus a file inventory of every module. Folders show where the code is; the graph shows what actually depends on what.
 
@@ -12,7 +12,7 @@
 | --- | --- | --- |
 | commands (CLI surface) | `src/commands` | 5 |
 | lib (core logic) | `src/lib` | 19 |
-| tests | `tests` | 46 |
+| tests | `tests` | 47 |
 
 ## Circular dependencies in `src`
 

context/map/artifact-3-contributors.md

@@ -2,7 +2,7 @@
 
 <!-- GENERATED by tools/repo-map/scan.mjs — do not hand-edit. Run `npm run repo-map`. -->
 
-> Scan of `7e3dab4` (2026-07-28) · window: last 12 months · engine: tools/repo-map
+> Scan of `c823d6d` (2026-07-28) · window: last 12 months · engine: tools/repo-map
 
 **Who to ask before you change something.** Bots and AI agents (Claude, Codex, Copilot, dependabot…) are filtered out — this is about humans who hold undocumented context. Answers "who has context on this area", not "who changed this line" (that is `git blame`).
 
@@ -20,8 +20,8 @@
 | Area | Ask (commits) |
 | --- | --- |
 | `src/lib` | “mkczarkowski” (29), psmyrdek (4) |
-| `context/changes` | “mkczarkowski” (28) |
 | `context/map` | Marcin Czarkowski (1) |
+| `context/changes` | “mkczarkowski” (28) |
 | `tests/e2e` | “mkczarkowski” (8), psmyrdek (3) |
 | `src/commands` | “mkczarkowski” (19), psmyrdek (4) |
 | `package.json` | “mkczarkowski” (7), Przemek Smyrdek (1) |

context/map/repo-map.md

@@ -2,7 +2,7 @@
 
 <!-- GENERATED by tools/repo-map/scan.mjs — do not hand-edit. Run `npm run repo-map`. -->
 
-> Scan of `7e3dab4` (2026-07-28) · window: last 12 months · engine: tools/repo-map
+> Scan of `c823d6d` (2026-07-28) · window: last 12 months · engine: tools/repo-map
 >
 > Operational map of the territory, built from cheap deterministic signals (git history + import graph), synthesized for decisions — not an essay. A new developer should know in ~15 min where things live, what is dangerous, and where to start. Refresh with `npm run repo-map`.
 
@@ -19,7 +19,7 @@
 graph TD
   commands["commands (CLI surface)<br/>5 files"]
   lib["lib (core logic)<br/>19 files"]
-  tests["tests<br/>46 files"]
+  tests["tests<br/>47 files"]
 ```
 
 ## 2. Terrain — where the project lives
@@ -29,11 +29,11 @@ Top active areas (full table in `artifact-1-territory.md`):
 | Area | Commits |
 | --- | --- |
 | `src/lib` | 75 |
+| `context/map` | 54 |
 | `context/changes` | 53 |
-| `context/map` | 50 |
 | `tests/e2e` | 32 |
 | `src/commands` | 30 |
-| `package.json` | 28 |
+| `package.json` | 29 |
 | `.github/workflows` | 19 |
 | `context/archive` | 19 |
 
@@ -61,22 +61,22 @@ _No high-risk zones flagged this run._
 | Area | Ask |
 | --- | --- |
 | `src/lib` | “mkczarkowski” (29), psmyrdek (4) |
-| `context/changes` | “mkczarkowski” (28) |
 | `context/map` | Marcin Czarkowski (1) |
+| `context/changes` | “mkczarkowski” (28) |
 | `tests/e2e` | “mkczarkowski” (8), psmyrdek (3) |
 | `src/commands` | “mkczarkowski” (19), psmyrdek (4) |
 | `package.json` | “mkczarkowski” (7), Przemek Smyrdek (1) |
 
 ## 6. First day — read these first
 
-- `package.json` — 28 commits in window
+- `package.json` — 29 commits in window
 - `src/commands/get.ts` — 16 commits in window
 - `.github/workflows/ci.yml` — 14 commits in window
 - `tests/get-command.test.ts` — 13 commits in window
 - `src/lib/writer.ts` — 12 commits in window
-- `context/map/artifact-1-territory.md` — 10 commits in window
-- `context/map/artifact-2-structure.md` — 10 commits in window
-- `context/map/artifact-3-contributors.md` — 10 commits in window
+- `context/map/artifact-1-territory.md` — 11 commits in window
+- `context/map/artifact-2-structure.md` — 11 commits in window
+- `context/map/artifact-3-contributors.md` — 11 commits in window
 
 ## 7. Limitations — what this map does NOT say
 

context/night-loop.md

@@ -0,0 +1,58 @@
+# Night loop runner
+
+`scripts/night-loop.mjs` runs bounded, local Codex iterations against
+`claude/context-loops-bootstrap`. It is intentionally narrower than the full
+loop menu: config-only lint ratchets that already pass. Characterization work
+is excluded until mutation proof can be executed mechanically; examples work
+is excluded because its executable config cannot safely run in an unattended
+outer gate.
+
+## Safety contract
+
+- Hard local-time deadline (default: the next `07:30`).
+- One agent and one isolated git worktree at a time.
+- Codex runs with `workspace-write`, `approval_policy=never`, and network
+  disabled. The runner—not the agent—performs authenticated GitHub operations.
+- Maximum 8 iterations and 800,000 reported tokens by default. A single Codex
+  turn can overshoot the remaining token budget because usage is reported only
+  after the turn; no repair or later iteration starts after the ceiling.
+- One base commit is pinned for the entire run. If the remote base advances,
+  the loop stops instead of comparing against a stale baseline.
+- Exact differential gate: lint cannot regress; passing tests cannot decrease;
+  the pinned set of 8 build-dependent failures must remain byte-for-byte the
+  same. Characterization work must increase the passing-test count.
+- At most two repair turns in the same Codex session. Persistent failure opens
+  an `automation` issue and preserves the failed worktree for diagnosis.
+- Successful iterations open one small PR each. The runner never merges PRs.
+- Agent and gate environments omit credential variables. Codex and outer gates
+  run without network; binary diffs and patches matching common secret formats
+  are rejected.
+
+The machine must remain awake and online. Saved Codex authentication and `gh`
+authentication must already work.
+
+## Review and dry run
+
+```bash
+bun test tests/night-loop.test.ts
+bun run night-loop -- --deadline 07:30 --dry-run
+```
+
+## Start an overnight run
+
+```bash
+nohup bun run night-loop -- --deadline 07:30 \
+  > .codex/night-loop-launcher.log 2>&1 &
+```
+
+State, command logs, final model outputs, and the final summary are written
+under `.codex/night-loop/<run-id>/`. The directory is ignored by git.
+
+Optional bounds:
+
+```bash
+bun run night-loop -- \
+  --deadline 2026-07-28T07:30:00+02:00 \
+  --max-iterations 4 \
+  --max-total-tokens 400000
+```

package.json

@@ -31,7 +31,8 @@
     "repo-map": "node tools/repo-map/scan.mjs",
     "repo-map:check": "node tools/repo-map/scan.mjs --check",
     "examples:catalog": "node tools/examples-catalog/build.mjs",
-    "examples:catalog:check": "node tools/examples-catalog/build.mjs --check"
+    "examples:catalog:check": "node tools/examples-catalog/build.mjs --check",
+    "night-loop": "node scripts/night-loop.mjs"
   },
   "dependencies": {
     "@clack/prompts": "0.9.1",

scripts/night-loop-lib.d.mts

@@ -0,0 +1,47 @@
+export interface LintSummary {
+  warnings: number;
+  errors: number;
+}
+
+export interface TestSummary {
+  pass: number;
+  fail: number;
+}
+
+export interface Baseline {
+  lint: LintSummary;
+  tests: TestSummary;
+  failedTests: string[];
+}
+
+export interface DiffStats {
+  files: number;
+  additions: number;
+  deletions: number;
+  binaryFiles: number;
+}
+
+export type LoopKind = "lint";
+
+export function resolveDeadline(now: Date, value?: string): Date;
+export function parseLintSummary(output: string): LintSummary;
+export function parseTestSummary(output: string): TestSummary;
+export function parseFailedTests(output: string): string[];
+export function parseCodexEvents(output: string): {
+  threadId?: string;
+  totalTokens: number;
+  completedTurns: number;
+};
+export function parseNumstat(output: string): DiffStats;
+export function validateScope(
+  kind: LoopKind,
+  paths: string[],
+  stats: DiffStats,
+): string[];
+export function validateDifferential(
+  kind: LoopKind,
+  baseline: Baseline,
+  candidate: Baseline,
+): string[];
+export function scanPatchForSecrets(patch: string): string[];
+export function validateLintRuleChange(baseText: string, candidateText: string): string[];