text
stringlengths
3
8.33k
repo
stringclasses
52 values
path
stringlengths
6
141
language
stringclasses
35 values
sha
stringlengths
64
64
chunk_index
int32
0
273
n_tokens
int32
1
896
--- phase: 02-foundation-configuration-sqlite-cas-runs-and-job-queue plan: "07" type: execute wave: 6 depends_on: [02-06] files_modified: - src/jobs/process-supervisor.ts - src/jobs/process-supervisor.test.ts - src/jobs/job-worker.ts - src/jobs/job-worker.test.ts - test/integration/process-lifecycle.test.ts a...
codex-hackathon
.planning/phases/02-foundation-configuration-sqlite-cas-runs-and-job-queue/02-07-PLAN.md
Markdown
98dcec87d0319479c3dd1931bee58059f7c1a02438a83b00fa6770e5731a8d7b
0
896
signal/timing/truncation, and unregister in finally. Keep platform-specific group handling capability-reported.</action> <verify><automated>bun x vitest run src/jobs/process-supervisor.test.ts</automated></verify> <done>The supervisor has one deterministic cleanup path and cannot accept command strings.</done> </ta...
codex-hackathon
.planning/phases/02-foundation-configuration-sqlite-cas-runs-and-job-queue/02-07-PLAN.md
Markdown
2a7b9362d263ae2c5d57483b15c2c2e3194f014814d34f62fa248dbd06ab1c80
1
667
--- phase: 02-foundation-configuration-sqlite-cas-runs-and-job-queue plan: "08" type: execute wave: 5 depends_on: [02-02, 02-04, 02-05] files_modified: - src/storage/usage.ts - src/storage/usage.test.ts - src/storage/gc.ts - src/storage/gc.test.ts - test/integration/gc-storage.test.ts autonomous: true require...
codex-hackathon
.planning/phases/02-foundation-configuration-sqlite-cas-runs-and-job-queue/02-08-PLAN.md
Markdown
33fe7ad6baf1cac6154a91e0437cca99c61f8a29a2e5a3ee65aacc04d2e11c06
0
896
.</action> <verify><automated>bun x vitest run src/storage/usage.test.ts src/storage/gc.test.ts</automated></verify> <done>Usage is deterministic and sweep cannot delete a referenced, new, changed, linked, or malformed object.</done> </task> <task type="auto" tdd="true" gate="REFACTOR"> <name>Task 3: REFACTOR — P...
codex-hackathon
.planning/phases/02-foundation-configuration-sqlite-cas-runs-and-job-queue/02-08-PLAN.md
Markdown
49f40b45365af7053dc0db0e3fb764dd76bdcaa673e01bb4a1a62939650ca7e7
1
693
--- phase: 02-foundation-configuration-sqlite-cas-runs-and-job-queue plan: "09" type: execute wave: 6 depends_on: [02-08] files_modified: - src/cli/gc.ts - src/cli/gc.test.ts - src/cli/main.ts - src/cli/main.test.ts - src/cli/command-tree.ts - src/cli/command-tree.test.ts - test/integration/gc-command.tes...
codex-hackathon
.planning/phases/02-foundation-configuration-sqlite-cas-runs-and-job-queue/02-09-PLAN.md
Markdown
0cd6b2bf0787b77d500e20c285881c36436242ac7b4f270409d2b3835cdaf877
0
896
leaf, mark only that leaf implemented, and inject a thin handler through `runCli`. Resolve the owned root/catalog through established dependencies, call Plan 08 with dry-run unless the flag is present, and return a stable envelope without importing the brownfield UI graph.</action> <verify><automated>bun x vitest run...
codex-hackathon
.planning/phases/02-foundation-configuration-sqlite-cas-runs-and-job-queue/02-09-PLAN.md
Markdown
9a400620c24731a13104cafc631a8663cc54550bb7e57e366c6cd473e63a9ea3
1
735
# Phase 2: Foundation: configuration, SQLite, CAS, runs, and job queue - Context **Gathered:** 2026-07-16 **Status:** Ready for planning **Mode:** Auto-generated (pure infrastructure phase) <domain> ## Phase Boundary Establish the contained, durable foundation used by every later MLX workflow: validated configuratio...
codex-hackathon
.planning/phases/02-foundation-configuration-sqlite-cas-runs-and-job-queue/02-CONTEXT.md
Markdown
3c4ae32ef0b1059c0b0a07bf01dd065516e824c50dc7c1264ed5fefb5971067b
0
590
# Deferred Items - `bun run check` remains red on pre-existing operator-owned files outside Plan 02-01, including `lib/discovery/validate/parse.test.ts`. The seven Plan 02-01 files pass a scoped Biome check, all 30 focused tests, and typecheck. No unrelated file was changed to repair the repository-wide baseline. - Pl...
codex-hackathon
.planning/phases/02-foundation-configuration-sqlite-cas-runs-and-job-queue/deferred-items.md
Markdown
3a475c93ec09f419258d1bb54af8dcb38b94ffd029e3c1a56fdabd446a7ca143
0
247
# Architecture Patterns **Domain:** Local-first personal coding dataset, benchmark, and model pipeline **Project:** MLX — the personal coding dataset and model pipeline **Researched:** 2026-07-15 **Overall confidence:** HIGH for product boundaries and build order; MEDIUM for dependency-specific implementation details ...
codex-hackathon
.planning/research/ARCHITECTURE.md
Markdown
a75eff6ef760256e441ecb97010a160804b2095c13cc9a72a1b488b00bc0c005
0
896
--+ +------------+--------------------+ | | +---------------+---------------+ | v +--------------+----------------+ | Downstream consumers | | Persona...
codex-hackathon
.planning/research/ARCHITECTURE.md
Markdown
f263a92abfeb47771010596a04332d192551cf58d78a166abf348ad940fc6d1a
1
896
definitions, exclusions, aggregation, population/as-of metadata | LLM inference, UI-specific formatting | Git, catalog, Parquet writer | | `packages/evidence` | Immutable evidence extraction, task grouping, accepted-state components, provenance | Dataset split assignment, training serialization | Git/metrics, CAS, Parq...
codex-hackathon
.planning/research/ARCHITECTURE.md
Markdown
d3a2d64b4cc81faf7312d9dd16504b0d60810f676309825713a54f64532b57ca
2
896
, benchmark | One row per `task_group_id`; computed before variant expansion and immutable within dataset version | | `DatasetReleaseManifest` | Dataset orchestrator after Python validation | Training, benchmark, Studio, publish | Immutable versioned manifest with snapshot, schema, split, dedup, privacy, file checksums...
codex-hackathon
.planning/research/ARCHITECTURE.md
Markdown
65f51892e97d58abfdaa914e9b6d37b702bf6c7e1fa5f0d718e938051844a09c
3
896
, execute bounded tools, capture patch/trace, remove and reconcile | | Immutable JSON manifests | Reproducibility and commit point for snapshots, runs, releases, experiments | High-volume analytics | Atomic write after referenced artifacts verify; never edit in place | SQLite should store CAS digests and Parquet paths...
codex-hackathon
.planning/research/ARCHITECTURE.md
Markdown
51ceee3e111c2dff90e419118dd0ccb295647810eaf6f9e8557bfa838c208d53
4
896
--------> blocked (dependency terminal failure) ``` `blocked` is derived or explicitly explained; it is not a silent terminal state. A run completes only after all required children are terminal and its manifest has been finalized. ### Claim and Lease Protocol 1. A worker opens a short write transaction, selects one...
codex-hackathon
.planning/research/ARCHITECTURE.md
Markdown
dbed179b973844ac02f2e8280668577a8a1859020c2ae63ba7630c8764abe6b2
5
896
| Contract tests and leakage corpus; private content blocked for unapproved endpoint | | Process/check execution | Command injection, environment leakage, orphan processes, destructive scripts | `check_id` -> fixed executable/argv template, no shell interpolation, cwd containment, minimal env, timeout/output cap, proce...
codex-hackathon
.planning/research/ARCHITECTURE.md
Markdown
9d8fce7df4fed6083a4a227e5b697bb15779f272541edf7d6202a07abcc8e6d4
6
896
acceptance tests before redirecting each command. 5. Relocate only synthetic, deterministic test assets to `fixtures/`; label old demo output `FIXTURE` or `REPLAY`, never `LIVE`. 6. Delete an old path only after its target replacement passes, no active entrypoint imports it, and its artifact disposition is recorded. 7....
codex-hackathon
.planning/research/ARCHITECTURE.md
Markdown
7ab637f6cd686a05eeffd0ff01500cf3b03beab8023293810aa178c907aacb9f
7
896
| Collision test cannot overwrite unrelated executable; user-facing identity audit; stable commands exist and report pass/fail/skip honestly; replacement matrix covers all legacy entrypoints/artifacts | None | | 2 | **Foundation: configuration, SQLite, CAS, runs, and job queue** | Typed config/root paths, storage packa...
codex-hackathon
.planning/research/ARCHITECTURE.md
Markdown
f598c57c207f32b073cf6186abd0ad1b1bd0123a68b1004bdd610d742cf0b7e6
8
896
contention, database growth, poor analytical scans, and unclear artifact integrity become coupled. **Instead:** SQLite stores state and references; CAS stores bytes; Parquet stores facts. ### Two Schema Authorities **What:** Hand-maintain unrelated TypeScript types, Python Features, and JSON schema definitions. **W...
codex-hackathon
.planning/research/ARCHITECTURE.md
Markdown
dc9b0f69f3ba56119c287d1f701ad65330f3f114f0ba34b267b314ae211b9f25
9
896
`Json()` and multi-config repository metadata with a real local load test. - **Phase 6:** Define the initial check registry and macOS process-group cleanup semantics before accepting repository execution. - **Phase 7:** Inspect installed MLX-LM help and model cards at execution time; no remembered flag set is an archit...
codex-hackathon
.planning/research/ARCHITECTURE.md
Markdown
842c3d2acd963b9ab738e50119572e4b10c1b0306e524d40b8ffb51c7e1c53b2
10
637
# Feature Landscape **Domain:** Local-first personal coding evidence, dataset, preference, benchmark, training, and presentation pipeline **Product:** MLX — the personal coding dataset and model pipeline **Researched:** 2026-07-15 **Overall confidence:** HIGH for required product capabilities; MEDIUM for current exter...
codex-hackathon
.planning/research/FEATURES.md
Markdown
82ba2ecc01007e4d51a15abcff6cc7818e5e12553a0d9c4a016dfdf265cec11b
0
896
dataset validation, benchmark smoke, Studio build, and capability-gated Apple Silicon checks provide repeatable acceptance rather than ad hoc demos. | Med | Script naming; fixture strategy | **Partial** — test/typecheck/check exist; required integration, Studio, dataset, benchmark, and local acceptance scripts are abse...
codex-hackathon
.planning/research/FEATURES.md
Markdown
3231aaaac4b29db1e21aaf714320f5ea5aab03bf43e294c78289159ba9860be2
1
896
pruned updates, immutable snapshot IDs, interruption recovery, and no content written into the application repository. | High | Explicit selection; jobs; contained paths | **Missing** | PS §§7, 8.2, 28.3, 29 | | Identity normalization with coverage reporting | Names, emails, GitHub noreply addresses, `.mailmap`, co-aut...
codex-hackathon
.planning/research/FEATURES.md
Markdown
82f47ade92f957d2c3180bce78135ee5e1760a8fee6f7c669277bebcc96a1cf1
2
896
local unless cross-repository evidence supports promotion. | High | Evidence; quality tiers; split policy | **Missing** | PS §17, §28.4; BS §§2.3, 3 | | Prompt-ready and machine-readable profile artifacts | Each snapshot emits `developer_profile.json` as the evidence graph and `DEVELOPER_PROFILE.md` as a concise profil...
codex-hackathon
.planning/research/FEATURES.md
Markdown
814535bad146728f636983cfcf2e2ced66ee1f586cd6e6980aaf15140e67f51c
3
896
29; DC §§4.10-4.11, 13-14; RR "Croissant metadata" | | Derived training and evaluation exports | Canonical data compiles deterministically to MLX-LM train/valid/test JSONL, TRL SFT/preference views, and LightEval-compatible data. Exports record source fingerprint, chat template, token estimates, truncation, and prefix-...
codex-hackathon
.planning/research/FEATURES.md
Markdown
2c6ad2e39afb5daab3ce6a8d42c6c14c226c1eb0c2539dbff0110dad771bde13
4
896
----------|---------------------| | Version-compatible MLX-LM preflight | The operator inspects installed `mlx_lm.lora --help`, model/chat-template/tool compatibility, disk and memory headroom, and records exact software/hardware versions rather than assuming remembered flags. | Med | Derived MLX-LM export; model acqui...
codex-hackathon
.planning/research/FEATURES.md
Markdown
6964b009880a5dfa8647e1567f10b4d3eb885d223b9344b3c08a6e0d8ac505e4
5
896
and datasets. Secret/PII/license/policy scans, redaction, exact upload preview, checksums/signatures, explicit `--push`/`--publish`, and confirmation precede any Hub or provider egress. | High | Catalog policy; evidence/dataset provenance | **Missing** — local model calls are local by default, but the complete scanner,...
codex-hackathon
.planning/research/FEATURES.md
Markdown
9a4d9d89899214ee4b8bd07d629eeb26229699aa0488248f34517c4d7f6ac32a
6
896
§§10, 27, 29; DC §§2, 14; BS §7 | | Honest local presentation of private evidence | A polished offline Studio can show real private work without uploading it, and its explicit data-origin labels prevent replay or fixtures from overstating event progress. | High | Studio; privacy; run manifests | **Missing** | PS §§3.2,...
codex-hackathon
.planning/research/FEATURES.md
Markdown
bb0c7753e0f6ad5051f4dba57e10e751c75f4837b9ce19535f390ddd02346991
7
896
metrics, traces, datasets, or adapters can leave the machine without informed review. | Local-private defaults, scanners, exact egress preview, explicit action, confirmation, and provider policy. | PS §§4.2, 26; DC §12 | | Random row splitting or sibling variants across splits | Near-identical examples and shared sourc...
codex-hackathon
.planning/research/FEATURES.md
Markdown
1cfcca9027d1831a1f0de73768004579889343b1a6bd45846db32f5937bf748a
8
896
| Giant root modules or UI-owned orchestration | Duplicated stage logic and hidden artifact coupling make recovery, testing, and migration unsafe. | Use explicit typed package boundaries and shared non-UI orchestration. | PS §6; CB "Anti-Patterns" | ## Feature Dependencies ```text Boundary 1 identity + migration base...
codex-hackathon
.planning/research/FEATURES.md
Markdown
19e4220bb0515377167a4f591488cbf1b0f1a52dadfc9e97b32c9364881c7d99
9
896
30; BS §10 | | Paired comparison | E4B base versus E4B base+profile with sample-level results, visible checks, and confidence intervals. | PS §30; BS §§6, 10 | | Live execution with fallback | One genuine live task execution plus deterministic replay fallback. | BS §10 | | Optional tuning result | Tuned E2B or E4B may ...
codex-hackathon
.planning/research/FEATURES.md
Markdown
1e9c66357a458dd00aa04113ba81c6004b07ca4a5f11a6f9ef2a27d0b9664299
10
896
and reinforced by dependency analysis. | | Event success floor | HIGH | Explicit in PS §30 and BS §10; kept separate from project completion. | | Brownfield status | HIGH | Cross-checked against the current architecture map, file inventory, package scripts, and targeted repository search. | | Hugging Face/MLX-LM ecosys...
codex-hackathon
.planning/research/FEATURES.md
Markdown
f4414e43cada1a5080bc288ca6bb2f2d71f06dbd77edb65472eb9b6cfced6dce
11
225
# Domain Pitfalls **Domain:** Local-first personal coding evidence, dataset, benchmark, and model pipeline **Project:** MLX — the personal coding dataset and model pipeline **Researched:** 2026-07-15 **Overall confidence:** HIGH for contract-specific risks; MEDIUM for dependency/runtime details that must be rechecked ...
codex-hackathon
.planning/research/PITFALLS.md
Markdown
990a3f31165e3e792451b3a697b1c4005c5773d4c4f19b2394740cc35817fdd0
0
896
scan. Never overwrite an unrelated executable without operator action. **Recovery and verification:** Remove only the link or binary installed by this project, restore the pre-install path entry if recorded, and rerun doctor from a clean shell. Verify both collision and no-collision fixtures and scan built artifacts, ...
codex-hackathon
.planning/research/PITFALLS.md
Markdown
e86bfd550a3df8f7e0d42784c81782f7817868d12f214bff1e27e6db433fe459
1
896
-coded to `main`; metric records omit source SHA and Git version/options; merge and non-merge counts do not reconcile; `--numstat` hyphens enter arithmetic; rename-heavy commits double count paths; `.mailmap` coverage and unmatched identities are hidden; blame output is labeled “all lines written.” **Prevention:** Use...
codex-hackathon
.planning/research/PITFALLS.md
Markdown
0444a5fb09bfaddec04bacf26755331ec7b12b015358fe2bee5a4537808243ab
2
896
worktrees; P7 model outputs; P8 final security audit. ### 8. Content addressing is implemented as hash-shaped filenames without integrity **What goes wrong:** CAS objects are written non-atomically, trusted by pathname without rehashing, collected while referenced, or hashed with ambiguous serialization. **Why it ha...
codex-hackathon
.planning/research/PITFALLS.md
Markdown
f8d37bf78462f06c3a6b2add2cfaf5e1bb62d3d96518126e1a35ea23cba4715e
3
896
cancellation and bounded retry categories; capture process groups and exit codes. **Recovery and verification:** Reap expired leases on startup, prove output identity before reuse, clean orphan processes/worktrees, reconcile job events with final manifests, and fault-inject crashes at claim, checkpoint, side-effect, c...
codex-hackathon
.planning/research/PITFALLS.md
Markdown
a872c565de278096330b7ba544421671e11583bbac77a83371473f5e80ea1e4f
4
896
retain base/head/parents and source refs, append new derivations, and require complete provenance before a record can exceed Q1. **Recovery and verification:** Re-extract from the pinned mirror snapshot, compare deterministic fields, supersede rather than mutate bad evidence, and demote any record whose complete linea...
codex-hackathon
.planning/research/PITFALLS.md
Markdown
0b707e095346acf56937ddf4d855e532ded187bccc1f26fa095676f46bd56ce6
5
896
define null/empty/timestamp/float behavior, and keep training exports as deterministic views. Schema changes require compatibility classification and release-version changes. **Recovery and verification:** Reject the release, migrate into a new immutable dataset version, validate every config/split and edge-case fixtu...
codex-hackathon
.planning/research/PITFALLS.md
Markdown
13c485009167a6e3da9a58edf0f1bf88ddc6bc36e057b48b63786a9a23c40f95
6
896
. **Warning signs:** Scans apply only to Parquet; logs include environment or subprocess command lines; raw traces are considered public after pseudonymizing repository IDs; `public_demo` shares source paths/text; cloud endpoints are configurable without egress classification; fixture bundles are not scanned. **Preve...
codex-hackathon
.planning/research/PITFALLS.md
Markdown
0403b614988d295ccd9694081e673b3c659fbbc857892d0df933afffe0d9ea32
7
896
the reference patch/future commit, writes outside the disposable worktree, invokes arbitrary shell through a check, inherits secrets, accesses the network, or mutates the mirror/shared cache. **Why it happens:** Tool names are fixed, but their host implementations accept raw paths, shell strings, ambient environment, ...
codex-hackathon
.planning/research/PITFALLS.md
Markdown
cc80229813e755d836cbc855f8a6924082efda1f60bfb25061b22538572f161b
8
896
## 25. Benchmark statistics turn noise into a win claim **What goes wrong:** Tiny differences on 20-50 tasks are shown without paired uncertainty, failed/missing tasks are omitted, aggregate composites hide execution regressions, or repeated comparisons are cherry-picked. **Why it happens:** A single percentage or pr...
codex-hackathon
.planning/research/PITFALLS.md
Markdown
920e87e6e5fc22a0bdc45d68728603ce1aceb125e9ae13e7a1506578a7a6a0d0
9
896
reported as Apple Silicon MLX acceptance **What goes wrong:** Mock subprocesses, Linux CI, fixture logs, or an E2B-only run are described as proof that E4B QLoRA, checkpoint/resume, adapter inference, memory, and throughput work on the target M4 Pro. **Why it happens:** Hardware runs are expensive and event pressure ...
codex-hackathon
.planning/research/PITFALLS.md
Markdown
8c3864a0a0593e4b591f622dad547859d29e4d53c6559f1e62faa5f8ddb9ae8e
10
896
verification:** Stop the server, inspect exposure window/logs, rotate any exposed credentials, and test network reachability from non-loopback interfaces before restart. **Owning phases:** P8 for Studio/API; P6/P7 for model endpoints. ### 31. Event success floor is upgraded into a project completion claim **What goe...
codex-hackathon
.planning/research/PITFALLS.md
Markdown
267a0989702e0521f104b8e8a629d12d293e60379a415b789aaf73ddf0aa979b
11
896
**Owning phases:** P1 baseline; P5 HF toolchain; P7 MLX-LM; P8 reproducibility audit. ### 4. LLM retry policy hides systematic semantic failures **What goes wrong:** Bounded semantic workers repeatedly regenerate until schema/leakage/quality checks pass, biasing the retained corpus while losing failure evidence. **W...
codex-hackathon
.planning/research/PITFALLS.md
Markdown
15245fca6b55fd08c9353af8a662acc3ae4c6cb3feb646cbc5fbe3d05c536d8e
12
896
percentages and abbreviated counts no longer sum or match downloadable data. **Prevention:** Preserve exact values in APIs/artifacts, state denominators, and round only at the final display layer with an “other/rounding” explanation. **Owning phases:** P3 and P8. ### 2. Pseudonymous repository IDs are unstable **Wh...
codex-hackathon
.planning/research/PITFALLS.md
Markdown
1b13d3b9f7326c36ae3e11cb5c68ddec8f2b0a155a2ca699bb72cc0ad6f05089
13
896
leased jobs, crash/process recovery | Upgrade/interruption tests; CAS closure/hash audit; lease race/fault injection; WAL/disk-full recovery; no writes outside roots | | P3 | Explicit authorization, complete paginated inventory, mirror recovery, identity coverage, exact Git metric definitions and exclusions | Synthetic...
codex-hackathon
.planning/research/PITFALLS.md
Markdown
4a4e7617fbd8c4f07ffe660ab37c44a3983a3914d34df5d63a8068b90664af41
14
896
structure: https://huggingface.co/docs/datasets/en/repository_structure - Hugging Face Datasets loading: https://huggingface.co/docs/datasets/en/package_reference/loading_methods - Hugging Face dataset cards and upload guidance: https://huggingface.co/docs/hub/en/datasets-cards and https://huggingface.co/docs/hub/en/da...
codex-hackathon
.planning/research/PITFALLS.md
Markdown
7730006132bb8afb158da3334ad97fc9980d8448c32f463ddf8be4d425eceab2
15
591
# Technology Stack **Project:** MLX — the personal coding dataset and model pipeline **Research type:** Subsequent/brownfield stack research **Researched:** 2026-07-15 **Overall confidence:** MEDIUM-HIGH ## Executive Recommendation Build the target as two deliberately separate planes: 1. **Bun/TypeScript control pl...
codex-hackathon
.planning/research/STACK.md
Markdown
c670afb21c71a8f7a5de9b91482ef81089203565c9879cd9ac4d1de2e6010ee2
0
896
one catalog owner/write connection and expose reads through the local API. See the mandatory WAL gate below. | MEDIUM | | Numbered SQL migrations | Repository-owned | Database evolution | Prefer explicit `migrations/NNNN_name.sql` plus a migration ledger and checksum over an ORM-generated mutable schema. Test fresh ins...
codex-hackathon
.planning/research/STACK.md
Markdown
2d051255a6114f0983f27c43dd651a88bff4256e4ff79adfe9cf6940649069b5
1
896
every shard schema and metadata. | HIGH | | `polars` | 1.42.1 | Lazy deterministic transforms and aggregations | Use `scan_parquet` and lazy expressions for evidence/metrics tables. Do not let Polars infer the canonical HF nested schema. | HIGH | | `duckdb` | 1.5.4 | Analytical SQL and cross-artifact audits | Use read-...
codex-hackathon
.planning/research/STACK.md
Markdown
f57935e67345f7f9321b0e0b4df79d3c34e5e4ee5ee3b3a69aa314585874dba7
2
896
runtime | Pin the upstream-paired minimum instead of silently accepting MLX 0.32.0. Move the pair together only after memory, checkpoint/resume, adapter load, and generation acceptance. | MEDIUM | | Transformers | Locked compatible 5.x (current 5.14.0) | Tokenizer/chat template/model metadata | Let uv lock an exact 5.x...
codex-hackathon
.planning/research/STACK.md
Markdown
79325b599934b726578ef6ec0846a79565d726783372eb7fb8e8663f13ba88e5
3
896
/config and scanner version; preserve finding class and redacted location. Never log matched secret values. | HIGH | | Presidio Analyzer | 2.2.363, optional | Local PII candidate detection | Combine with deterministic email/path/account rules and operator review. Treat it as a candidate detector, not proof that a relea...
codex-hackathon
.planning/research/STACK.md
Markdown
b78e7bd5f99864dcb5c9aa2fea0c4fe3776c09a2d14563c08efde71f508334a5
4
896
version").get())' # Locked installs bun install --frozen-lockfile uv --version uv python find 3.13 uv lock --check uv sync --locked # Python package matrix uv run python -c 'import importlib.metadata as m; print({p:m.version(p) for p in ["datasets","pyarrow","polars","duckdb","huggingface-hub","jsonschema"]})' # Ins...
codex-hackathon
.planning/research/STACK.md
Markdown
a929678b123787e6647d1be9455c0a95f00467d3430441b050ae5362ba831504
5
896
a public application platform. Add a framework only if routing complexity becomes real. | | Studio | React + Vite | Next.js/Electron | No SSR, cloud deployment, or desktop packaging requirement justifies the extra runtime. Browser plus loopback service is smaller and easier to audit. | | Repository metadata | `gh api` ...
codex-hackathon
.planning/research/STACK.md
Markdown
3cc0886ebe8598edee1d0975858e61d687e0219b3ca7b32308bbbadbd32b1652
6
896
| Replace | Implement the fixed host tools from the contract and an allowlisted check registry. Preserve useful schema validation/fuzz patterns. | | Existing `data/` runtime paths | Replace | Move state under `MLX_HOME`; add CAS, manifests, retention, and tests proving no writes escape configured roots. | | Training sh...
codex-hackathon
.planning/research/STACK.md
Markdown
39e9d0430255d1d9944ef17dd95158bdec852315195565da11f25c50063259c5
7
896
-client conflict. Value at personal-corpus scale remains unproven, so defer. | | PII scanning | MEDIUM | Gitleaks is strong for secrets; general PII detection needs policy-specific deterministic rules, optional Presidio, and operator review. | ## Roadmap Consequences The stack implies this order: 1. Pin identity/too...
codex-hackathon
.planning/research/STACK.md
Markdown
75716cc8733381a119f2fe38099246702144f40b4e07ec5688c33fe169c5fa1a
8
896
-hub/1.23.0/) - [Hugging Face Hub CLI reference](https://huggingface.co/docs/huggingface_hub/en/package_reference/cli) - [TRL dataset formats](https://huggingface.co/docs/trl/en/dataset_formats) - [MLCommons Croissant](https://mlcommons.org/working-groups/data/croissant/) - [DataTrove 0.9.0 package metadata](https://py...
codex-hackathon
.planning/research/STACK.md
Markdown
3c8e3cd174e121e3eb3761ae143631bf2f80bbf05953da41edd0f57c65f2998f
9
489
# Project Research Summary **Project:** MLX — the personal coding dataset and model pipeline **Domain:** Local-first personal coding evidence, dataset, benchmark, training, and presentation pipeline **Researched:** 2026-07-15 **Confidence:** HIGH for product boundaries and architecture; MEDIUM-HIGH for the pinned impl...
codex-hackathon
.planning/research/SUMMARY.md
Markdown
e46e9a53a50977226ed273d0e233a0119f36ed6e4d9dce05c1b334e4ca3b5fd9
0
896
*`bun:sqlite` + numbered SQL migrations:** operational catalog for repositories, identities, snapshots, jobs, leases, runs, and artifact references; require foreign keys, WAL, bounded busy handling, short transactions, and the runtime ownership gate. - **SHA-256 CAS:** immutable blobs, patches, traces, reports, and lar...
codex-hackathon
.planning/research/SUMMARY.md
Markdown
67b8898f594178bb0f1d0316736e51e296787c8eea79c2fcd069cc0e8b579bb7
1
896
versus base-plus-profile versus tuned versus tuned-plus-profile comparisons that isolate prompting gain from tuning gain. - End-to-end lineage from authorized repository snapshot through evidence, split, dataset, export, model, benchmark, and Studio result. - A useful dataset/profile/benchmark product even when no trai...
codex-hackathon
.planning/research/SUMMARY.md
Markdown
f010f7845b2abc2950b693fc86d137f0c5d8776f33d2043cc7e45879c0a5c74a
2
896
for merges, reverts, renames, binaries, and blame. 4. **Filesystem, catalog, or execution escape** — centralize containment, reject traversal/symlink/archive attacks, keep catalog transactions short, use fenced job attempts, and execute only reviewed checks in disposable worktrees. 5. **Leaky or irreproducible datasets...
codex-hackathon
.planning/research/SUMMARY.md
Markdown
5028394762e66848ddca2b1032de03bbdafadf6a6910320be986be53c033357c
3
896
/vendor inflation, and blame mislabeled as total authorship. **Exit boundary:** Unselected repositories are never cloned; pagination and permission gaps are visible; interrupted mirror updates recover; fixture metrics reconcile exactly with documented populations, exclusions, and timestamps. ### Phase 4: Evidence Extr...
codex-hackathon
.planning/research/SUMMARY.md
Markdown
c968d17a48f8cc2f7159dcc095a67972f123119114dddfcdd811f78cd21239e7
4
896
-LM exports, real M4 Pro acceptance, experiment lineage, and separation of prompting versus tuning gain. **Avoids:** Remembered MLX-LM flags, incompatible templates/masks, unsupervised intermediate tool actions, silent dependency drift, unpaired comparisons, and capability skips reported as passes. **Exit boundary:** S...
codex-hackathon
.planning/research/SUMMARY.md
Markdown
22241b418a430e2dc81f9342844d92ec38de883e4be13394c8fd791f6fd25272
5
896
. - **Phase 3:** Freeze current GitHub GraphQL/REST queries, authentication scopes, pagination, rate-limit behavior, Git option semantics, and fixed tool versions against official interfaces and fixtures. - **Phase 5:** Resolve and lock the complete Python dependency graph; re-run explicit nested Features, multi-config...
codex-hackathon
.planning/research/SUMMARY.md
Markdown
d143bd9c6896321b3c52a27dfda55dc44869f69e3cbbf66ad97af90862eb22b6
6
896
`](../../docs/MLX_PROJECT_SPEC.md) — product principles, architecture, storage, required features, eight independent roadmap boundaries, full definition of done, and event success floor. - [`docs/MLX_DATASET_CONTRACT.md`](../../docs/MLX_DATASET_CONTRACT.md) — canonical schemas, configs, task groups, quality, deduplicat...
codex-hackathon
.planning/research/SUMMARY.md
Markdown
c5526dbc26f938535b06205cda41cca94329e630efcf11590a3e0bb16d84a7a9
7
670
{ "tools": [ { "type": "function", "function": { "name": "supabase_rls_policy_template", "description": "Emit a Supabase RLS policy DDL template for a given table, role, and operation.", "parameters": { "type": "object", "properties": { "tableNam...
codex-hackathon
data/adapter-tools.fallback.json
JSON
6388fa1d6213e7d790babbfdb9f82341918f0219b7945d490b53bc184dc621ad
0
896
] }, "value": {} }, "required": [ "column", "op", "value" ], "additionalProperties": false } } }, "required": [ ...
codex-hackathon
data/adapter-tools.fallback.json
JSON
8643ad54d3801f50f4ab86b0d5f70b4152aa59071d49c42706ed6ddcb279534d
1
896
type": "string" }, "userId": { "type": "string" }, "filename": { "type": "string" } }, "required": [ "bucket", "userId", "filename" ], "additionalProperties...
codex-hackathon
data/adapter-tools.fallback.json
JSON
1c479c8cc93fa81aea6f105df75ff64e4ab61507142d9c8a4a20374d914c415e
2
896
lowercase letter and contain only a-z, 0-9, hyphens, underscores" } }, { "userPrompt": "Validate empty edge function name.", "call": { "name": "supabase_edge_function_name_validator", "arguments": { "name": "" ...
codex-hackathon
data/adapter-tools.fallback.json
JSON
5a3ca6559c284a9743eeeeed9aa473fc45ae2168f4649b1cff8e23f4668be144
3
896
[^:\\/]+)(?::(\\d+))?\\/(.+)$/.exec(url); if (!m) return { host: '', port: 5432, database: '', user: '' }; return { host: m[3] || '', port: m[4] ? Number(m[4]) : 5432, database: m[5] || '', user: m[1] || '' }; }", "requiresNetwork": false, "trajectories": [ { "userPrompt": "Parse p...
codex-hackathon
data/adapter-tools.fallback.json
JSON
54414c5ccbbcfb88f9ce454863f751c82fa9f5e2d034c88f878fecbb567f3fb3
4
896
[i + 2] !== '=') out += String.fromCharCode((n >> 8) & 255); if (b[i + 3] !== '=') out += String.fromCharCode(n & 255); } try { var obj = JSON.parse(out); return { sub: typeof obj.sub === 'string' ? obj.sub : null, role: typeof obj.role === 'string' ? obj.role : null, exp: typeof obj.exp === 'number' ? obj.exp : null }...
codex-hackathon
data/adapter-tools.fallback.json
JSON
6b35e736eca8cbd37f46ec97f5fa2cd443230f773186220cbbfeb25e9b13aca7
5
896
.", "call": { "name": "supabase_migration_filename", "arguments": { "timestampMs": 1700000000000, "description": "Add users table" } }, "result": { "filename": "20231114221320_add_users_table.sql"...
codex-hackathon
data/adapter-tools.fallback.json
JSON
80d9dcbfb01fe4f210e4fbb81ed0a4c2e4f0e6cbf0ed0d1716f37887bb8c6ba1
6
325
{ "chunks": [ { "id": "llms.txt#0000", "source": "llms", "text": "Row Level Security (RLS) lets you restrict which rows a given user can read or modify. You enable RLS on a table by running ALTER TABLE your_table ENABLE ROW LEVEL SECURITY. Once enabled, all queries are filtered through policies ...
codex-hackathon
data/corpus.json
JSON
de1628ab2a258acf4fa58640148ae2b943ffd9c35ef378b31fb9720073bca4e0
0
896
let you deploy server-side TypeScript on Deno at the edge. Each function exports a default handler receiving a Request and returning a Response. You deploy with supabase functions deploy my-function. Edge Functions access your database through the service role key, bypassing RLS for admin operations. You set secrets us...
codex-hackathon
data/corpus.json
JSON
5621056f72b98fb7125c8e926fcacdf33e16554260161a666b25e7230e56160d
2
402
# MLX PersonalBench Specification ## 1. Purpose MLX PersonalBench measures whether a model has learned transferable engineering behavior from the MLX dataset rather than memorizing source files. The benchmark compares models on the same tasks, tools, context budget, step budget, checks, and hardware policy. --- ##...
codex-hackathon
docs/MLX_BENCHMARK_SPEC.md
Markdown
7d3c20ee5d3d7b5d369e7811a5e2d8ea7297e090a43256b9e95cd0375db7dd99
0
896
version, - tool and prompt versions, - hardware/software versions, - seed and sampling parameters, - check registry, - full observable trace, - score breakdown. Run a contamination/leakage audit before publishing results. --- ## 8. Hugging Face and LightEval integration - Store benchmark tasks in the dataset `bench...
codex-hackathon
docs/MLX_BENCHMARK_SPEC.md
Markdown
2832ca9834e9d34bbdab9395b0b0df4611cb6f9c4ad0fac2e595744fe817b4fa
1
274
# MLX Hugging Face Dataset Contract **Purpose:** Define the canonical schemas, configs, quality gates, split rules, and release artifacts for the MLX personal coding dataset. --- ## 1. Design goals The dataset must be: - useful through Hugging Face `datasets` without custom loading code, - queryable as Parquet/Arr...
codex-hackathon
docs/MLX_DATASET_CONTRACT.md
Markdown
86b0796590050bfd9e48d7e7e4eaddfeb9d1d102d62932a243f51a99c334a00a
0
896
. ### 4.6 `preference` ```text id: string task_group_id: string prompt: Json chosen: Json rejected: Json preference_source: string chosen_verification: Json rejected_verification: Json metadata: Json ``` ### 4.7 `benchmark` ```text task_id: string task_group_id: string suite: string repository_id: string base_sha: ...
codex-hackathon
docs/MLX_DATASET_CONTRACT.md
Markdown
b2fd9a55b94ba28a1074a2df3b4408cc8968982f28b7340b942b3762def098c0
1
896
-change state and grounded task information. Target may be: - accepted patch, - final assistant response, - or a verified alternative patch. Reject when: - target diff appears in prompt, - task cannot be inferred without future information, - change is generated/vendor/format-only, - patch is incoherently large, - ...
codex-hackathon
docs/MLX_DATASET_CONTRACT.md
Markdown
b1fa72179b7b8b2a0e6d29a7013be21e9e6000a7a3172b48dc8402bbd29ba34c
2
896
only private-hub metadata-only public-samples-approved excluded ``` The dataset card must describe source rights and limitations. Do not imply that private source code is redistributable. Before upload: - secret scan, - PII scan, - source visibility check, - license/policy check, - exact upload preview, - explicit c...
codex-hackathon
docs/MLX_DATASET_CONTRACT.md
Markdown
a614fd656ba863ff20ea9bf37bbfdf10afc0b066dfc766c0b27677cfa75a94f7
3
342
# MLX Event Fast Track — July 16, 2026 This is an honest presentation slice, not a replacement for the full project roadmap. ## Goal Demonstrate that MLX can turn real GitHub history into: 1. accurate engineering metrics, 2. an evidence-backed coding preference profile, 3. a Hugging Face-native dataset, 4. and a he...
codex-hackathon
docs/MLX_EVENT_FAST_TRACK.md
Markdown
817e5e6ada20fd52669e2520a4473fe55b7205e21c959febe73f3a990a4d12c1
0
505
# MLX Presentation Runbook ## Stage narrative ### Act 1 — My GitHub DNA Show: - selected repository count, - public/private/internal mix without private names, - authored commits, - additions/deletions/churn, - current SLOC, - surviving attributed lines, - languages over time, - active days. Explain that each metr...
codex-hackathon
docs/MLX_PRESENTATION_RUNBOOK.md
Markdown
8cd5dff9543509dc4e35ba47793cc39381acdcd1e49b5b11a022de542ef3084a
0
490
# MLX — Project and Acceptance Specification **Status:** Authoritative product contract **Target machine:** Apple M4 Pro, 24 GB unified memory **Primary event date:** July 16, 2026 **Product:** MLX — the personal coding dataset and model pipeline **CLI:** `mlx` **Local state:** `~/.mlx` or `MLX_HOME` --- #...
codex-hackathon
docs/MLX_PROJECT_SPEC.md
Markdown
648b2e605a716d7a9d3fda5218592e3f4cc77ecb87e41c75f7a46ca59bc13ff5
0
896
not collect hidden model chain-of-thought. - Do not execute arbitrary model-generated shell commands. - Do not claim the tuned model is better without paired, held-out evidence. - Do not optimize exclusively for reproducing whitespace or formatting. - Do not require the iPhone deployment path from the earlier project. ...
codex-hackathon
docs/MLX_PROJECT_SPEC.md
Markdown
5c590c3532376b4b8a813ed3cc90d21e836fa61616b7ca18a5b89a010ba5db1f
1
896
merge authored commits, - merge commits, - co-authored commits, - commits by repository/language/year/month, - active days, - longest activity streak, - time-of-day and day-of-week distribution, - median and percentile patch size, - files changed, - additions, deletions, net change, and churn. Definitions: ```text ne...
codex-hackathon
docs/MLX_PROJECT_SPEC.md
Markdown
709918e190ad2b327f4388f5704c99bbe7e4d7afc14946b68ed5e3ab2d171783
2
896
- **Intent Reconstructor** — converts issue/PR/commit evidence into a clean task statement without leaking the target implementation. - **Context Selector** — chooses a minimal sufficient pre-change file/symbol context. - **Preference Miner** — proposes preference claims with support and counter-evidence. - **Trajector...
codex-hackathon
docs/MLX_PROJECT_SPEC.md
Markdown
2d167ef71a04d8a96544992e574437445128091bda7a54bf46f91916e52d9476
3
896
selected training repositories, reserve the latest meaningful history segment for evaluation. Compute the cutoff per repository and record it. ### 16.3 Whole-repository holdout Reserve complete repositories stratified by language, framework, size, and importance. These repositories must contribute no source content o...
codex-hackathon
docs/MLX_PROJECT_SPEC.md
Markdown
6fbd456540494eda8c0cfd0bd09e66aba54dc60a6b95643bd76e51d8c0231b5e
4
896
must not be sent to cloud APIs without explicit sanitization and approval. Primary benchmark categories: - repository task execution, - localization/navigation, - style/preference choice, - whole-repository generalization, - temporal/future generalization. Report task-level outputs, not only aggregate scores. --- ...
codex-hackathon
docs/MLX_PROJECT_SPEC.md
Markdown
fbf13acb956816c28c2c312ce212e88b9ec7b6a20fdf495d606e36ab3d6817ce
5
896
and export, - PII scanning and configurable redaction, - repository visibility propagation, - path traversal and symlink escape tests, - shell argument safety, - archive extraction safety, - no raw credentials in logs, - no private content in fixture bundles, - explicit egress policy for teacher or judge providers, - l...
codex-hackathon
docs/MLX_PROJECT_SPEC.md
Markdown
0ab4f90877ccbeba40309945eef5d47c01f965cc36e7cb8eb2b831523de5d564
6
816
# MLX Research Rationale This document records the reasoning behind the dataset-first architecture. Implementation agents must re-check installed versions and current official documentation before pinning dependencies or command-line flags. ## Hugging Face Datasets - Hugging Face Datasets is Arrow-backed and designe...
codex-hackathon
docs/MLX_RESEARCH_RATIONALE.md
Markdown
eff45876c0360bcdce62eac8bf538f946ae6bd8dcc4003eabd84a512aa084aa4
0
896
not a success criterion. ## Research sources reviewed on July 15, 2026 Primary documentation: - Hugging Face Datasets documentation: https://huggingface.co/docs/datasets/en/index - Hugging Face dataset upload decision guide: https://huggingface.co/docs/hub/en/datasets-upload-guide-llm - Hugging Face dataset cards: h...
codex-hackathon
docs/MLX_RESEARCH_RATIONALE.md
Markdown
da581141761255027b0693f953855145333c4ed592e491b5ca0a0aad612beff0
1
392
{ "schemaVersion": "1", "executableName": "mlx", "expectedHuman": { "owned": "OWNED", "collision": "COLLISION", "notFound": "NOT FOUND" }, "repeatCalls": 2, "parallelCalls": 6, "scenarios": [ "owned-first", "foreign-first-owned-shadowed", "foreign-only", "none", "duplicate-...
codex-hackathon
fixtures/phase-1/doctor/candidates.json
JSON
15bbfdb1f2bf4b297b26812a54047112109807ac1075c20ed58cfebd1506ab12
0
178
{ "schemaVersion": "1", "canonicalPhrase": "MLX — the personal coding dataset and model pipeline", "appleDistinction": "This product is distinct from Apple's MLX project and is not affiliated with or endorsed by Apple.", "requiredCategories": ["documentation", "help", "package", "screenshots"], "sources": [ { ...
codex-hackathon
fixtures/phase-1/identity/scoped-tree.json
JSON
aae2d5616b86d94654629ecf420a354636bcf85b0619b58cb44b2de5567fd8a3
0
896
shell/path.txt", "kind": "file", "content": "/operator/unrelated/bin:/usr/bin\n" }, { "path": "shell/profile.sh", "kind": "file", "content": "export PATH=/operator/unrelated/bin:$PATH\n" } ] } }
codex-hackathon
fixtures/phase-1/identity/scoped-tree.json
JSON
c9b1df940d63c4006d2536da8889647d649100251dd6fb0042ffcfac561ef448
1
81
{ "source": "FIXTURE", "tools": [ { "type": "function", "function": { "name": "supabase_rls_policy_template", "description": "Synthetic schema-only fixture for deterministic validation tests.", "parameters": { "type": "object", "properties": { ...
codex-hackathon
fixtures/phase-1/validation/adapter-tools.json
JSON
a44b6b017e9986be1181f8b3c19f3625e1cefea4afb80e5b1417ec0ff7e10650
0
182
// AdapterLoaderView.swift — enumerates Documents/adapters/* directories that // carry a LoRA adapter pair (adapter_config.json + adapters.safetensors), // offers Load / Unload buttons, shows lastSwapMs and a deterministic probe diff. // // Mount into ios/_upstream/Applications/LLMEval/ via Xcode "Add Files to LLMEval"...
codex-hackathon
ios/SpecialistApp/AdapterLoaderView.swift
Swift
0b89692252fc00c08766167fa029ef8dca7890741c8fbb4613c4fbe000c1e506
0
896
path guard fm.fileExists(atPath: path, isDirectory: &isDir), isDir.boolValue else { return false } let w = url.appending(component: "adapters.safetensors").path let c = url.appending(component: "adapter_config.json").path return fm.fileExists(atPath: w) && fm.fileExists(a...
codex-hackathon
ios/SpecialistApp/AdapterLoaderView.swift
Swift
3aa631c191a568811d66774169e6888b14853f9b6cdf48772eabf967195c5622
1
100
import Foundation private struct AdapterToolManifest: Decodable { let tools: [AdapterToolRecord] } private struct AdapterToolRecord: Decodable { let function: AdapterToolFunction let meta: AdapterToolMeta } private struct AdapterToolFunction: Decodable { let name: String let description: String ...
codex-hackathon
ios/SpecialistApp/AdapterToolsLoader.swift
Swift
c1f190337ae5472fdbe72d486cb459374a7d15ca074361ec98232f45a80a81fa
0
274
import SwiftUI enum ChatRole { case user case assistant case toolCall } struct ChatMessage: Identifiable { let id = UUID() let role: ChatRole let content: String var toolName: String? = nil var toolArgs: String? = nil var toolResult: String? = nil } struct ChatView: View { @En...
codex-hackathon
ios/SpecialistApp/ChatView.swift
Swift
88989c699d3b5ab2eebc8b6994409a93a5c4c3d865b115ea1a63454e172e029b
0
896
.calls.enumerated() { let resultText = index < result.results.count ? result.results[index] : "" let argsData = try? JSONSerialization.data(withJSONObject: call.args, options: [.prettyPrinted]) let argsText = argsData.flatMap { String(data: $0, encoding: .utf8) } ?? "{}" ...
codex-hackathon
ios/SpecialistApp/ChatView.swift
Swift
0d4b4b301fbf5975f7929525c8a97f60826263786a92c0a2f527365b9433acea
1
141