Skip to content

stevesolun/ctx / docs

skill-router

Flagged

Repo-aware recommendation manager for ctx. Scans the active repository, identifies stack and workflow signals, recommends a capped set of skills, agents, and MCP servers, and unloads helpers that no longer match the current work after user confirmation. Harnesses are recommended by the custom-model onboarding flow or loop adapters and then attach to the same recommendation layer.

FileKindSize
SKILL.mdskill md9.3 KB
assets/javascripts/catalog.jsscript2.5 KB
assets/javascripts/repo-stats-refresh.jsscript3.2 KB
backup-hook-install.mdmarkdown9.2 KB
catalog.mdmarkdown10.3 KB
ctx-fit/ARCHITECTURE_CANDIDATES.mdmarkdown12.9 KB
ctx-fit/ARCHITECTURE_GAP_ANALYSIS.mdmarkdown10.7 KB
ctx-fit/CTX_AUDIT.mdmarkdown11.6 KB
ctx-fit/CURRENT_STATE.mdmarkdown16.3 KB
ctx-fit/DECISIONS.mdmarkdown15.7 KB
ctx-fit/GOOD_FIRST_ISSUES.mdmarkdown20.6 KB
ctx-fit/IMPLEMENTATION_TRACKER.mdmarkdown697 B
ctx-fit/MAP.mdmarkdown7.8 KB
ctx-fit/PRD.mdmarkdown9.7 KB
ctx-fit/STATE.mdmarkdown60.9 KB
ctx-fit/milestones/M1_SPIKE_FRICTION_LOG.mdmarkdown5.5 KB
ctx-fit/milestones/MILESTONE_00_AUDIT.mdmarkdown5.1 KB
ctx-fit/planned_tasks.mdmarkdown17.7 KB
dashboard.mdmarkdown24.7 KB
enterprise-readiness-review.mdmarkdown8.7 KB
entity-onboarding.mdmarkdown18.0 KB
harness/attaching-to-hosts.mdmarkdown16.2 KB
harness/clean-host-contract.mdmarkdown4.5 KB
harness/loopflow-adapter-demo.mdmarkdown11.1 KB
huggingface-publish.mdmarkdown3.3 KB
index.mdmarkdown22.4 KB
knowledge-graph.mdmarkdown22.7 KB
marketplace-registry.mdmarkdown4.0 KB
memory-anchor.mdmarkdown2.7 KB
plans/ctx-agent-reviewer-workbench.mdmarkdown12.4 KB
plans/unified-capability-engine.mdmarkdown59.0 KB
qa/dashboard-user-story-status.csvother26.0 KB
qa/feature-user-story-status.csvother92.5 KB
research/context-engineering-guidance-audit.mdmarkdown10.4 KB
services/macos/com.claude.backup.watchdog.plistother1.7 KB
services/systemd/claude-backup-watchdog.serviceother1.3 KB
skill-lifecycle-and-dashboard.mdmarkdown6.2 KB
skill-quality-install.mdmarkdown6.2 KB
skill-router/index.mdmarkdown2.5 KB
skill-stack-matrix.mdmarkdown5.5 KB
skills-health.mdmarkdown3.5 KB
stack-signatures.mdmarkdown6.2 KB
telemetry.mdmarkdown24.7 KB
threat-model.mdmarkdown15.0 KB
toolbox/behavior-miner.mdmarkdown2.6 KB
toolbox/configuration.mdmarkdown4.0 KB
toolbox/council-runner.mdmarkdown2.4 KB
toolbox/hooks.mdmarkdown2.6 KB
toolbox/index.mdmarkdown3.0 KB
toolbox/intent-interview.mdmarkdown3.0 KB
toolbox/starters.mdmarkdown3.0 KB
toolbox/templates/docs-review.jsonconfig599 B
toolbox/templates/fresh-repo-init.jsonconfig533 B
toolbox/templates/refactor-safety.jsonconfig634 B
toolbox/templates/security-sweep.jsonconfig636 B
toolbox/templates/ship-it.jsonconfig676 B
toolbox/verdicts.mdmarkdown4.0 KB

ctx-fit/STATE.md

View on GitHub
# CTX Fit — Live Working State

> **This is the canonical operational checkpoint for current CTX Fit work.**
>
> Read `AGENTS.md` and `docs/ctx-fit/DECISIONS.md` first. Then read this file
> before resuming a long-running task. The root `STATE.md` is frozen history
> from the superseded unified-capability-engine goal and is not live state.
>
> This file records where the work is. Code, tests, and accepted decisions
> remain the source of truth for what the product does and where it is going.

## Checkpoint

- Updated: 2026-08-21 (Asia/Jerusalem)
- Active goal: retire Git LFS safely after CTX Fit 1.0.21
- Phase: repository and Mac-local cleanup complete; remote LFS purge pending
- Release decision: **1.0.21 REMAINS RELEASED; LFS MIGRATION MERGED; REMOTE PURGE PENDING**
- Branch: `main`
- Release commit: `38a33f8784e2bf408430a98fed81206c2cf39d00`
- Release tag object: `a7b8e78559fda1d44dca844393458272071ae89b`
- LFS migration PR: `https://github.com/stevesolun/ctx/pull/275`
- Cleanup checkpoint PR: `https://github.com/stevesolun/ctx/pull/276`
- Follow-up scope at checkpoint:
  - ask GitHub Support to purge the historical remote LFS objects
  - add repository/environment protection rules as defense in depth
  - choose a retention policy before removing Codex task transcripts
  - user-owned and out of scope: `.scratch/`
- Parallel execution: complete. Independent product, security, spend,
  activation, packaging, release, Linux, SBOM, and recovery reviewers accepted
  the shipped tree. Exact-main Tests run `31914958343`, CodeQL run
  `31914958371`, and Hugging Face sync `31914958347` are green. Production
  publish run `31915534546` completed successfully, including reproducible
  package build, package and graph provenance, CycloneDX attestation, release
  assets, and PyPI Trusted Publishing.
- LFS migration execution: three parallel lanes completed repository resolver,
  workflow migration, and independent storage/identity audit. Merged `main`
  removes the two tracked archive pointers, LFS hooks/rules/fallbacks, and
  obsolete park/prune tooling; strict release-manifest hydration replaces them.
  Focused integrated evidence is 326 passed plus 113 reviewer-adjacent tests;
  full non-integration passed 8,812 tests with 5 skips; the authoritative PR
  preflight passed all 21 lanes. The final committed clean-checkout gate passed
  all 12 lanes after selectively hydrating the runtime catalog before unit
  tests. Independent final audit accepted the repaired tree with no P0-P2
  findings. PR #275 merged as `b62b78d7` after 19 successful checks and one
  intentionally skipped matrix placeholder. A detached checkout of that exact
  merge, with no Git LFS executable or local archives, hydrated and deeply
  validated the 405,434,548-byte release pair. Local cleanup removed
  1,968,180,750 bytes of release-backed LFS cache objects, 197,656,624 bytes of
  stale temporary packs, and the untracked 405,434,548-byte working pair: at
  least 2,571,271,922 bytes (2.395 GiB) total. A later Mac-wide audit found and
  removed a second 16-object, 4,534,205,612-byte LFS cache in the `no-mistakes`
  bare mirror, then restarted and verified the daemon. Safe cache and clean
  worktree cleanup reduced the Data volume's rounded used space from 280 GiB to
  262 GiB. No remote LFS object has been purged yet.

## Product destination

CTX Fit must find the cheapest capability configuration that reliably completes
representative work in the user's repository, using repository-native
verification rather than an agent's self-report. It must present evidence, keep
"current setup" as a valid result when that claim is actually supported, and
produce a reviewable change that reproduces the winning configuration.

The current release is limited by ADR-007 to capability configuration within a
single harness. It must not claim to compare Codex, Claude Code, or other
harnesses when it did not run that experiment.

## Definition of done for 1.0.21

All of the following must be true before tagging or publishing:

1. A production trial can edit only its throwaway workspace with the intended
   tools, and an untrusted repository cannot read ambient secrets or mutate the
   host outside that workspace.
2. The evaluated agent cannot change the verification judge or any file outside
   the task's explicit editable set and still earn a verified result.
3. No public API or CLI path can spend without an executable plan, an explicit
   budget, and user authorization after a pre-spend preview.
4. Baseline, candidate, dry-run, recommendation, and apply/PR output describe
   the configuration that is actually evaluated. Applying a winner reproduces
   it rather than writing internal IDs only.
5. Incomplete or one-sided experiments return no verdict. CTX-imposed budget
   truncation is inconclusive and remains auditable.
6. Representative task derivation works for the supported language set, or the
   release surface states a narrower, truthful support contract.
7. Focused tests, static checks, the fast gate, PR preflight, package smoke, and
   the exact release commit's required CI checks are green on the final tree.
8. Public docs, changelog, package metadata, and GitHub release notes describe
   1.0.21 truthfully. The release tag points to the exact reviewed green commit.

## Work status

| Workstream | State | Owner | Current evidence / next condition |
| --- | --- | --- | --- |
| Merge the reviewed CTX Fit base | Complete | Coordinator | PR #266 merged; this hardening branch starts at `bd36bbea` |
| Independent product review | Complete | Product reviewer | Verdict: do not release; 4 P0 and 7 P1 findings recorded below |
| Production agent editing surface | Complete, independently accepted | Coordinator + independent reviewer | Production trials expose only a workspace-rooted filesystem MCP through the shared sandbox, use a scrubbed environment, deliver exact candidate material, and now refuse missing harness dependencies before trial setup; the integrated 508-test Fit suite and Linux/provider refreeze are green |
| Repository sandbox and secret isolation | Complete, independently accepted | Sandbox writer + coordinator + independent reviewer | Exact executable/symlink paths cannot expose sibling trees, trusted runtime subtrees remain usable, repository setup/verification stays network-disabled, and provider authority remains separate. Final macOS refreeze passed 84 focused tests plus real child/installed-CTX compatibility and static checks with no P0-P2 findings |
| Verification-judge integrity | Complete, independently accepted | Coordinator + architecture reviewer | The forgeable Python witness was removed. ADR-016 defines the repository command as an explicit, non-adversarial trust boundary; exact Python/JS/TS/Go/Rust/Make commands run unchanged, verification writes are confined to one trial workspace, and the assumption appears before spend and in every result. Independent review found no P0-P2 |
| Spend authorization and preview | Complete, independently accepted | Coordinator + independent reviewer | Immutable digest-bound plans, human pre-spend preview, JSON plan-only behavior, strict simulator identity, exact caps, and honest observed over-cap accounting passed 106 focused tests plus independent adversarial review |
| Fair campaign completion | Complete, independently accepted | Coordinator + independent reviewer | Exact candidate/task/trial/floor identity, no partial verdicts, strict numeric and simulation handling, and full-precision selection have no remaining reviewer findings |
| Truthful current baseline | Complete, independently accepted | Coordinator + independent reviewer | Simple installed skills are exact content-addressed material; invalid/complex skills and unreproducible agent/MCP/tool configurations abstain. Baseline drift is checked before authorization, before spend, and after the campaign; 222 focused tests and static checks passed in independent refreeze |
| Reproducible apply/PR result | Complete, independently accepted | Apply writer + independent reviewer | Sidecar-only `.ctx/fit-configuration.json`; immutable exact materials, instruction preimage checks, CAS, symlink refusal, transactional rollback, and PR staging passed 91 focused tests plus Ruff/mypy and independent review |
| Multi-language task derivation | Complete, aggregate green | Multi-language writer + coordinator | Python, JavaScript, TypeScript, Go, and Rust paired source/test history contract; 446-test Fit aggregate passed |
| Doctor/runtime truth | Complete, independently accepted | Doctor writer + coordinator + independent reviewer | Live selection and credential forwarding are bound to the exact selected model; unused or mismatched keys cannot authorize live execution, exact-model pricing and invalid sidecars fail closed, and independent refreeze passed 156 focused tests plus static/read-only checks |
| Applied winner activation | Complete, independently accepted | Activation writer + coordinator + independent reviewer | Strict repository-root loader, nested-sidecar refusal, hash/model validation, one-use exact context, subdirectory activation, and explicit model-conflict handling passed independent refreeze as part of the 222-check baseline/activation lane |
| ADR-015 stop attribution | Complete, independently accepted | Coordinator + spend/fairness reviewer | Structured stop reason/log fields flow provider → live runner → serialized result, budget-capped trials are inconclusive, and the accepted spend/fairness lane plus 446-test Fit aggregate are green |
| Release metadata and publish guard | Complete, independently accepted | Metadata writer + publish writer/reviewer + coordinator | 1.0.21 metadata and changelog are current; exact-main/exact-successful-Tests production guard and changelog-backed notes have no P0/P1 review findings; P2 credential/doc hardening is integrated |
| Final verification and release | Complete, publicly verified | Coordinator + release/SBOM reviewers | PR #271 merged as release commit `38a33f8784e2bf408430a98fed81206c2cf39d00`; exact-main Tests, CodeQL, and Hugging Face sync succeeded. Annotated tag `v1.0.21` peels to that commit. Publish run `31915534546` completed every build, attestation, release-asset, and PyPI job. Public wheel/SBOM digests, Sigstore/Rekor attestations, clean installation, `pip check`, version output, and a minimal `ctx fit --json` repository profile were independently verified. |
| Retire Git LFS | Repository and Mac-local cleanup complete; remote purge pending GitHub Support | Coordinator + manifest/workflow writers + independent auditor | 45 historical objects total 12.131 GiB because each compressed archive revision is stored whole. Current useful pair is 405,434,548 bytes and is independently preserved and attested in release v1.0.21. PR #275 merged as `b62b78d7` after 19 successful checks. A clean post-merge checkout hydrated and deeply validated both release assets without Git LFS. The repository and `no-mistakes` mirror have no LFS objects or local LFS configuration; the development-root scan found no other LFS object store. Main and mirror integrity checks pass, `no-mistakes` is running, and `.scratch/` is untouched. The two cleanup passes reclaimed about 20 GiB locally; GitHub still retains and bills the historical remote objects until Support purges them. |
| Local storage hygiene | Reproducible CTX residue removed; user history preserved | Coordinator | The superseded A/B benchmark root fell from about 16.6 GiB to 440 MiB; only the dirty 14-file evaluator worktree remains. Repository `.gate` fell from 4.6 GiB to 136 KiB after old A/B runs, caches, environments, and a redundant checkout were removed. Its two-line uncommitted change is retained as a checked recovery patch. Stale Codex plugin staging freed another 76.8 MiB. Codex task transcripts (20.3 GiB active; 3.44 GiB archived) and the live logs database were inventoried but not deleted without a retention decision. Rounded Data-volume use is now 241 GiB. |

## Open release blockers

### P0 — release-stopping

None. All definition-of-done conditions were closed before publication.

### P1 — must resolve or explicitly de-scope before release

None. The CycloneDX PURL and deterministic content-derived RFC-4122 UUIDv5
serial repairs passed the production generator, strict closure validator, and
pinned attestation action in publish run `31915534546`.

The release deliberately discloses that qualification did not include a paid
provider call. The required Ubuntu lane proved Bubblewrap, Node, `npx`, the
optional harness, and zero-spend driver construction without invoking a model.
This is an evidence limit, not a claim the release makes.

External release settings remain a P2 operational risk: observed `main` and
the `pypi` environment have no server-side protection rules. The workflow now
fails closed unless the tag is the exact current `main` head with a successful
exact-SHA Tests run, but repository settings should still add reviewer/tag
protection after this release.

## Verification ledger

Evidence is valid only for the tree named in the row. Any edit to a covered
surface makes that row stale for release purposes.

| Tree / date | Check | Result | Release use |
| --- | --- | --- | --- |
| Pre-hardening tree, 2026-08-13 | `src/tests/fit` | 307 passed | Baseline only; stale after current edits |
| Pre-hardening tree, 2026-08-13 | package/surface/clean-host focused contracts | 62 passed | Baseline only; stale after current edits |
| First ADR-015 edit, 2026-08-13 | focused budget-stop tests | Passed | Partial behavior evidence only |
| First ADR-015 edit, 2026-08-13 | Ruff on changed ADR-015 files | Passed | Partial static evidence only |
| Current hardening tree, 2026-08-13 | focused judge integrity, budget-cap, log-retention, and ambient-secret selector | 6 passed | Local behavior evidence; aggregate and independent review still required |
| Current hardening tree, 2026-08-13 | Python compile + Ruff on sandbox/live-runner/execution surfaces | Passed | Local static evidence; provider integration is not complete |
| Current hardening tree, 2026-08-13 | direct sandbox adversarial tests | 3 passed | Proves this macOS host denies sibling write and ambient-temp read while allowing workspace writes |
| Current hardening tree, 2026-08-13 | provider boundary/translation suite | 20 passed | Shared boundary invocation, least-authority environment, tool surface, and structured result translation |
| Current hardening tree, 2026-08-13 | environment reuse integration + focused security selector | 7 passed | Dependency setup network split and workspace re-aiming are green on this host |
| Spend/fairness writer tree, 2026-08-13 | owned focused tests | 89 passed | Writer evidence only; independent review pending |
| Multi-language writer tree, 2026-08-13 | task derivation focused tests | 23 passed | Writer evidence only; aggregate pending |
| Exact-apply draft, 2026-08-13 | focused apply/candidate tests | 63 passed | Insufficient: independent reviewer rejected after 4 adversarial preservation/correctness probes failed |
| Spend/fairness draft, 2026-08-13 | focused reviewed files | 73 passed, 16 failed on shared candidate-fixture drift | Independent reviewer rejected with six adversarial classes; repair active |
| Later exact-apply draft, 2026-08-13 | focused apply/candidate tests + static checks | 87 passed; Ruff, format, mypy, diff check passed | Independent reviewer still rejected AGENTS mutation; sidecar-only repair active |
| Accepted sidecar-only apply, 2026-08-13 | focused apply/candidate tests + static checks + independent refreeze | 91 passed; Ruff, format, mypy, diff check passed; reviewer accepted | Valid lane evidence; aggregate still required |
| Integrated exact baseline/activation tree, 2026-08-13 | candidate/profile/experiment/provider/live/activation/harness CLI selector | 282 passed; source static checks passed before later spend edits | Strong local integration evidence; stale after spend changes and independent security refreeze active |
| Current spend integration tree, 2026-08-13 | execution/recommend/experiment/budget/Fit CLI selector | 100 passed; Ruff, format, mypy, diff check passed | Fresh writer/coordinator evidence; independent final refreeze active |
| Accepted spend/fairness tree, 2026-08-13 | execution/recommend/experiment/budget/Fit CLI selector + adversarial probes | 106 passed; Ruff, format, mypy, diff check passed; reviewer accepted | Accepted semantic lane evidence |
| Current verifier-witness tree, 2026-08-13 | `src/tests/fit/test_live_runner.py` + source static | 33 passed; Ruff, format, mypy, diff check passed | Writer/coordinator evidence; independent refreeze active |
| Rejected verifier-witness tree, 2026-08-14 | independent adversarial probes + focused regression selector | Existing 33 tests/static green, but 2 new attacks earn `verified`; selector is 2 passed/2 failed | Release-stopping red evidence; proves the previous witness is not an authority boundary |
| Current integrated Fit tree, 2026-08-13 | full `src/tests/fit` suite | 440 passed in 80.12s | Fresh integrated behavior evidence; later code edits invalidate covered surfaces |
| Truthful-dimensions repair, 2026-08-14 | profile + dry-run focused behavior/static | 17 passed; Ruff, format, mypy passed | The experiment now claims only skill-capability variation; aggregate still required |
| ADR-016 verifier-boundary repair, 2026-08-14 | live runner + plan/result/disclosure focused behavior | 58 passed after one prose-regression correction; live runner alone 32 passed; Ruff/format/mypy/diff passed | Exact native commands and workspace-only write boundary implemented; independent refreeze active |
| Accepted ADR-016 tree, 2026-08-14 | independent verifier/profile/task/sandbox/release-surface refreeze | 186 focused checks passed; real editable-install and campaign-reuse paths passed; Ruff/format/mypy/diff passed | Independent reviewer accepted with no P0-P2 |
| Current full non-integration tree, 2026-08-14 | repository-wide parallel pytest excluding browser/integration | 8,706 passed, 5 skipped, 1 tracker-attribution failure | Implementation evidence green; the sole governance failure named the two new Fit modules and was repaired immediately afterward |
| Tracker-repaired tree, 2026-08-14 | feature/bug/dashboard/toolbox tracker contracts | 36 passed | Canonical FIT-001 now attributes `applied_configuration.py` and `sandbox.py`; repository-wide pytest rerun still required |
| Tracker-repaired pre-refreeze tree, 2026-08-14 | repository-wide parallel pytest excluding browser/integration | 8,707 passed, 5 skipped | Full behavior evidence was green before the independently found baseline and read-boundary P0 repairs; stale for the final release tree |
| First-use baseline red/green slice, 2026-08-14 | exact installed repository skill appears in baseline material/context | Failed against empty-control implementation, then passed after exact materialization; candidate module 33 passed | Focused writer evidence only; unsafe/unreproducible layouts, aggregate, static, and independent refreeze remain |
| Applied-model activation repair, 2026-08-14 | applied/profile CLI selectors + static | 32 applied tests and 34 combined checks passed; Ruff, format, mypy passed | Writer evidence only; aggregate and independent refreeze remain |
| Accepted baseline/activation tree, 2026-08-14 | candidate/applied/provider/experiment/budget/apply/CLI selectors + static | 222 passed; Ruff, format, mypy, diff check passed; reviewer accepted | Exact current baseline, drift guards, repository-root activation, model binding, and safe-read availability have no P0/P1 findings |
| Current integration tree, 2026-08-14 | baseline/activation/apply/experiment/discovery/live-runner/sandbox/provider selector | 364 passed | Fresh integrated behavior evidence after the baseline and sandbox repairs |
| Current integration tree, 2026-08-14 | full `src/tests/fit` suite | 478 passed in 69.95s | Fresh Fit behavior evidence; final release still requires repository-wide and static gates |
| Accepted doctor/runtime-truth tree, 2026-08-14 | model-aware CLI/doctor/provider/budget/applied/experiment/profile tests + static/read-only probe | 156 passed; Ruff, format, mypy, diff check passed; reviewer accepted | Exact selected-model credential and pricing truth has no remaining material finding |
| Post credential/process repair tree, 2026-08-14 | critical sandbox/live/provider/doctor/budget/activation/experiment selector + static | 196 passed; Ruff, format, mypy, diff check passed | Fresh coordinator evidence, superseded for sandbox release use by the later exact-path refreeze finding |
| Post credential/process repair tree, 2026-08-14 | full `src/tests/fit` suite | 497 passed in 79.89s | Fresh integration evidence, but sandbox exact-path repair will require rerun |
| Accepted exact-path sandbox tree, 2026-08-14 | sandbox/live-runner/provider focused behavior + real macOS compatibility + static | 17 sandbox, 41 live-runner, and 26 provider tests passed; Ruff, format, mypy, diff check passed; reviewer accepted | Exact-file/runtime-subtree authority split and provider separation have no P0-P2 findings; Linux remains structural and no paid model call occurred |
| Current settled Fit tree, 2026-08-14 | full `src/tests/fit` plus public surface truth | 522 passed in 71.20s | Fresh settled behavior evidence; repository-wide, packaging, and committed-history gates remain |
| Base release audit, 2026-08-13 | release-contract suite | 201 passed | Baseline only; docs/code edits require rerun |
| Base `bd36bbea`, 2026-08-13 | reproducible wheel/sdist, manifest, and Twine checks | Passed twice | Baseline only; final-tree artifacts will have different hashes |
| Base `bd36bbea`, 2026-08-13 | GitHub Tests run `31703885499` | Completed successfully | Proves the merged base only; final tag SHA needs a fresh green run |
| Final uncommitted tree, 2026-08-14 | CI-shaped full non-integration suite | 8,761 passed, 5 skipped, 15 deprecation warnings in 262.33s | Green exact-tree behavior evidence. A prior parallel attempt had one deterministic-bridge request-count miss; the isolated test passed three times and this complete rerun passed |
| Final uncommitted tree, 2026-08-14 | Ruff check/format, mypy, strict MkDocs, trackers/release/package surfaces, repo stats, diff integrity | Ruff green across `src hooks scripts`; 615 files formatted; mypy green on 585 files; MkDocs strict green; 135 focused contracts passed; stats and diff checks green | Required uncommitted static/documentation/release evidence complete |
| Commit `0264cede`, 2026-08-14 | `scripts/no_mistakes_run.sh fast --allow-dirty` | All 11 lanes passed in 356.89s; committed-head-only report | Valid for `0264cede`; superseded after the CodeQL repair is amended |
| Commit `0264cede`, 2026-08-14 | `python scripts/ci_preflight.py --profile pr` from a clean detached worktree | All 19 lanes passed; 8,761 tests, 5 skipped, 92.15% coverage; reproducible artifacts and Twine green | Valid for `0264cede`; superseded after the CodeQL repair is amended |
| PR #267 / commit `0264cede`, 2026-08-14 | GitHub PR checks | Product/build/clean-host/static/docs/CodeQL-Python lanes green; aggregate CodeQL rejected one high world-readable-manifest alert | Release-stopping remote evidence; local repair active |
| CodeQL permission repair working tree, 2026-08-14 | exact new-manifest mode regression, full apply/Fit/surface suites, docs/stats, Ruff/format/mypy, independent probes | Regression failed at `0644`; 64 apply tests, 523 Fit/surface tests, docs/stats, and static checks passed with `0600`; reviewer verified create/modify/rollback under `umask 000` and accepted with no P0-P2 | Accepted repair evidence; new committed and remote gates required |
| Commit `2cc8667d`, 2026-08-14 | committed fast gate + clean detached PR preflight | All 11 fast lanes and all 19 preflight lanes passed; 8,762 tests, 5 skipped, 92.15% coverage; wheel `9986b8c6...`, sdist `099825fd...`, Twine green | Exact macOS/local release evidence; remote Linux still required |
| PR #267 / commit `2cc8667d`, 2026-08-14 | GitHub CodeQL and Tests | CodeQL aggregate plus 15 specialized checks passed; `unit-linux` failed 8 cases (2 missing Bubblewrap, 6 missing optional harness metadata/pricing), causing aggregate CI failure | Release-stopping Linux environment-contract evidence; parallel repair active |
| Reviewed Linux/provider remediation tree, 2026-08-14 | full Fit + Linux/provider/CLI + no-LiteLLM + workflow/CI/docs contracts | 508 Fit passed; 146 targeted passed; no-LiteLLM 72 passed/16 expected skips; workflow/CI 137 passed; docs tracker 36 passed; Ruff/format/mypy/YAML/embedded-Python/diff green | Independent integration reviewer accepted with no P0-P2; remote Ubuntu lane remains the authoritative Linux execution evidence |
| Commit `10e47d37`, 2026-08-14 | committed fast gate + clean detached PR preflight | All 11 fast lanes and all 19 preflight lanes passed; 8,769 tests, 5 skipped, 92.16% coverage; wheel `03cdaac7...`, sdist `1b488e9e...`, Twine green | Exact implementation commit evidence; this state-only follow-up needs proportional revalidation before push |
| Commit `6ec9d9b2`, 2026-08-14 | proportional clean release/docs/package revalidation | 180 release/workflow/docs tests, strict MkDocs, stats, reproducible build, and Twine passed; wheel `4d3e5e3c...`, sdist `798d9582...` | Exact pushed candidate evidence before remote Ubuntu execution |
| PR #267 / commit `6ec9d9b2`, 2026-08-15 | required Ubuntu live-prerequisite lane | 3 positive sandbox checks failed before child start with `bwrap: loopback: Failed RTM_NEWADDR`; 7 negative checks were not valid denial evidence because the child never started | Release-stopping remote evidence; directly drove the AppArmor profile, child-start sentinel, and operational-preflight repair |
| Accepted Ubuntu repair tree, 2026-08-15 | sandbox/provider/doctor/workflow focused + full Fit + independent refreeze + static/docs | Coordinator: 64 focused and 513 full Fit passed. Independent reviewer: 513 Fit, 99 focused, 75 adjacent, docs/tracker 36; Ruff, format, mypy, workflow parsing, strict MkDocs, stats, and diff checks all passed; no P0-P2 | Accepted uncommitted evidence; committed gates and exact-SHA remote Ubuntu run remain |
| Commit `c4aaa22e`, 2026-08-15 | committed fast gate + clean detached PR preflight | All 11 fast lanes and all 19 preflight lanes passed; 8,774 tests, 5 skipped, 92.16% coverage; wheel `e9770d99...`, sdist `a0d86319...`, Twine green | Exact local release evidence; remote Ubuntu still required |
| PR #267 / commit `c4aaa22e`, 2026-08-15 | required Ubuntu live-prerequisite lane | AppArmor profile loaded and every child started; 8 passed, while a private-root sibling write and private POSIX-shm creation contradicted the old denial assertions | Release-stopping policy/evidence mismatch; directly drove private-root remount and same-name shm isolation proof |
| Linux private-root repair tree, 2026-08-15 | sandbox/provider/doctor/workflow focused + static | 64 passed; Ruff, format, mypy, and diff check passed; independent review found no P0/P1 | Fresh local evidence; committed gates and exact-SHA Ubuntu remain |
| Release commit `38a33f87`, 2026-08-16 | local committed fast gate and clean PR preflight | 8,783 passed, 5 skipped, 92.16% coverage; all 11 fast lanes and all 19 preflight lanes passed; reproducible wheel/sdist and Twine green | Final local release evidence |
| PR #271 / head `6a409de9`, 2026-08-16 | GitHub Tests and CodeQL | Tests run `31914512225` and CodeQL run `31914512301` succeeded; unit-linux passed 8,680 with 50 skips and 91.32% coverage | Accepted merge evidence |
| Exact release commit `38a33f87`, 2026-08-16 | canonical main CI and graph distribution | Tests `31914958343`, CodeQL `31914958371`, and Hugging Face sync `31914958347` succeeded | Exact tag provenance evidence |
| Annotated tag `v1.0.21`, 2026-08-16 | production publish and public smoke | Publish `31915534546` succeeded; GitHub release, PyPI wheel/sdist, SBOM, package/graph/SBOM attestations, clean install, `pip check`, version, and minimal-repository Fit profile verified | Release complete |
| LFS migration working tree, 2026-08-21 | manifest/workflow/local-consumer focused integration | 326 passed with one expected tar deprecation warning; adjacent manifest/graph-validation/review suite 113 passed; writer lanes separately reported 137+57 and 165 passed with Ruff/format/mypy/YAML checks green | Strong uncommitted evidence; full gates and independent final-tree review remain |
| Accepted LFS migration working tree, 2026-08-21 | full non-integration, static/docs/package gates, independent adversarial refreeze, and authoritative PR preflight | 8,812 passed, 5 skipped in the full non-integration run; repaired surface 172 passed and independent refreeze 133 passed; Ruff, format, mypy, strict MkDocs, trackers, stats, clean-host, deep graph, browser, reproducible wheel/sdist, and Twine passed; all 21 PR-preflight lanes passed; reviewer accepted with no P0-P2 | Valid final-tree implementation evidence. One parallel-only deterministic-bridge miss in the first preflight attempt passed three isolated reruns and the complete retry; commit and committed-history gate remain |
| First committed LFS migration gate, commit `9e0fb368`, 2026-08-21 | isolated committed-head local-fast | Static, graph, docs, telemetry, similarity, browser, package, clean-host, canary, contract, and policy lanes passed; unit lane failed 123 benchmark/holdout fixtures because its clean worktree correctly lacked the now-untracked runtime archive | Valid red evidence: clean CI test jobs must selectively hydrate the 110,283,462-byte runtime catalog before pytest. Failing-first workflow/local contracts added; repair rerun pending |
| Repaired committed LFS migration, commit `72d1ad73`, 2026-08-21 | complete 12-lane committed-head local-fast, split only to avoid repeating an identical five-minute unit lane | Selective runtime hydration followed by the isolated unit lane passed 8,806 tests with 5 skips and 92.16% coverage. The other 11 exact-commit lanes passed static, policy, canary, contract, clean-host, docs, full graph hydration/deep validation, telemetry, similarity, browser security, reproducible package, and Twine checks | Valid complete committed-history evidence for the implementation commit; state-only checkpoint follow-up requires proportional docs/stats validation |
| PR #275 head `a7f49bc4`, 2026-08-21 | GitHub Tests and CodeQL | 15 specialized Tests checks passed while unit and graph jobs were still running; CodeQL analysis completed but security aggregation opened alert #23 (`py/overly-permissive-file`, high) because manifest refresh explicitly changed its atomic temp file to world-readable `0644` | Valid red evidence. Failing-first owner-only regression passes at `0600`; 74 focused/static checks passed and independent reviewer verified atomic replacement, rollback, cleanup, and supported-platform behavior with no P0-P2. Repair push/recheck pending |
| PR #275 final head `52ddcba6`, 2026-08-21 | GitHub Tests and CodeQL | 19 checks succeeded and one matrix placeholder was intentionally skipped; unit-linux, graph artifact, package smoke, clean-host, similarity, browser, telemetry, CodeQL, and required aggregation were green | Accepted merge evidence; permission repair and complete LFS migration are remotely verified |
| Merged `main` `b62b78d7`, 2026-08-21 | detached clean-host release hydration and deep graph validation | Without Git LFS or local graph archives, hydrated runtime SHA-256 `d4a39836...` (110,283,462 bytes) and full SHA-256 `2ce8a945...` (295,151,086 bytes); validated 79,958 nodes, 1,778,069 edges, and all catalog member counts | Exact post-merge proof that release assets replace the LFS operational dependency |
| Local checkout after merge, 2026-08-21 | LFS/config/garbage cleanup and repository integrity | No `.git/lfs`, LFS config, LFS-tracked path, or Git garbage remains; `git fsck --no-dangling` passed. `.scratch/` remains untouched | Local retirement complete; remote billed-object deletion remains external |

No paid real-provider evaluation has been authorized or run during this
hardening pass. A simulated or injected driver is not evidence of live provider
quality. Any paid canary requires a separately stated budget and consent.
TestPyPI does not yet contain a `claude-ctx` project, so that optional staging
path remains unproven. Production PyPI Trusted Publishing succeeded.

## Execution and review loops

### Resume loop

1. Read `AGENTS.md`, `docs/ctx-fit/DECISIONS.md`, and this file.
2. Inspect the actual branch, commit, working tree, and active agents. Preserve
   user-owned files and do not assume this checkpoint is newer than Git.
3. Reconcile this file with code and test evidence. If they disagree, code,
   tests, and accepted decisions win; update this file immediately.
4. Take the highest-severity unblocked item. Do not duplicate an active writer's
   files.
5. Update this checkpoint after a meaningful repair, new blocker, verification
   result, handoff, commit, merge, or release action.

### Implementation loop

1. Reproduce the defect with the smallest failing test or deterministic probe.
2. Make the smallest coherent repair within one owned surface.
3. Run the focused test and applicable static checks on the actual changed
   tree.
4. Send the repair through an independent semantic review when it affects
   security, spend, recommendation validity, apply behavior, or release flow.
5. Record the result and residual risk here; a writer's self-report alone is
   never verification.

### Parallel dispatch rules

- The coordinator owns scope, dependency decisions, integration, state, and the
  final release decision.
- Writers receive disjoint files and bounded outcomes. Shared files have one
  owner at a time.
- Reviewers do not silently repair the code they are judging.
- Expensive gates wait until focused failures are closed, so time and compute
  are not wasted proving a known-bad tree is bad.

### Release loop

1. Close every P0 and resolve or truthfully de-scope every P1.
2. Rerun focused tests, then Fit, full non-integration, Ruff, and mypy.
3. Commit the reviewed tree and run the fast and PR-preflight gates against that
   exact history.
4. Open and review the hardening PR. Merge only with required checks green.
5. Verify the resulting `main` SHA and package metadata; tag that exact SHA.
6. Publish through the tag workflow, verify PyPI and GitHub release artifacts,
   and record digests and links here.

## Immediate next actions

1. Ask GitHub Support to purge all 45 historical LFS objects (12.131 GiB).
   Pointer removal alone does not release GitHub's billed remote storage.
2. Recheck billed storage after Support confirms purge and after the next
   billing-cycle reset.
3. Add `main`, tag, and `pypi` environment protection rules as defense in depth;
   the shipped workflow already enforces exact-main and exact-successful-Tests
   provenance.
4. Decide whether to permanently remove Codex task history. Archived transcripts
   older than 30 days account for 3.43 GiB; active transcripts older than 90
   days account for 0.455 GiB. The current task and recent history must remain.

### GitHub Support handoff

Support contact: `https://support.github.com/contact`

**Subject:** Purge retired Git LFS objects for `stevesolun/ctx`

**Request:**

> Repository: `https://github.com/stevesolun/ctx`
>
> We merged the no-history-rewrite Git LFS retirement in PR #275, merge commit
> `b62b78d704ec54b3ff43e333982eef713a33c8df`. The repository no longer tracks,
> downloads, uploads, or needs any Git LFS object. Current graph artifacts are
> exact, attested GitHub Release assets and were verified from a clean checkout.
>
> Please purge every Git LFS object associated with this repository: 45 unique
> objects totaling 13,025,281,486 bytes (12.131 GiB), and confirm when the
> repository's billed LFS storage is cleared. We explicitly do not want the
> repository deleted/recreated and do not need a history rewrite.
>
> The Enterprise Cloud repository LFS DELETE endpoint returned HTTP 404 for this
> personal GitHub.com repository with both the legacy API version (request ID
> `C027:354872:36D09:4F64A:6A88AD3E`) and current API version `2026-03-10`
> (request ID `C52A:166CDA:57D73:880EA:6A88B8ED`). The latter used the active
> repository administrator's OAuth token with `repo` scope. The safe
> self-service route is therefore not available.

## Checkpoint log

- 2026-08-21: Retried the documented `DELETE /repos/stevesolun/ctx/lfs`
  endpoint with GitHub REST API version `2026-03-10`. The active `gh` OAuth
  token has `repo` scope, the authenticated user owns the repository, and the
  repository permission response reports `admin: true`; GitHub nevertheless
  returned HTTP 404 with request ID `C52A:166CDA:57D73:880EA:6A88B8ED`.
  Nothing changed remotely. GitHub's public removal contract still leaves only
  repository deletion/recreation or Support-assisted purging; deletion was not
  attempted because it would destroy repository identity and is outside this
  safe cleanup. Support purge remains the sole remote blocker.
- 2026-08-21: Removed the remaining reproducible local CTX experiment residue
  after reconciling it with the CTX Fit pivot. Deleted three clean detached
  benchmark worktrees whose commits are ancestors of `main`, all old A/B graph
  catalogs, raw runs, tool homes, and inactive sparse Lima/Colima disks. Kept
  the dirty `codex/fix-evaluator-status-superset` worktree and all 14 of its
  changes. `.ctx-benchmark` fell from about 16.6 GiB to 440 MiB. Removed the
  4.6 GiB repository `.gate` A/B corpus and redundant checkout after preserving
  its two-line `PYTEST_ADDOPTS` change in
  `.gate/recovery/ctx-ab-disable-pytest-cache.patch` (SHA-256
  `74c4c7b982e669b097049de9af49ab195d07a169fc3bee5fcca7629dc8d311e3`);
  `.gate` is now 136 KiB. Removed 76.8 MiB of stale Codex plugin staging while
  keeping the current bundled marketplace. Audited 20.3 GiB of active Codex
  transcripts and 3.44 GiB of archived transcripts; no transcript or live
  database was deleted because that would remove user history. Rounded Data
  volume use fell from 262 GiB to 241 GiB during this pass. Local Git LFS
  remains empty. GitHub's 12.131 GiB remote purge is blocked only on Support
  sign-in and submission.
- 2026-08-21: Completed a safe Mac-wide follow-up cleanup. Found 16 duplicate
  CTX LFS payloads totaling 4,534,205,612 bytes (4.223 GiB) in the active
  `no-mistakes` bare mirror. Stopped the daemon, removed only its LFS payloads
  and local LFS settings, restarted it, and verified the mirror and main Git
  object databases. Removed the obsolete generated CTX A/B catalog cache,
  Python download cache, abandoned Codex runtime install, explicitly broken SDK
  cache, Homebrew-confirmed obsolete files, and old gate sandbox/tool caches.
  Removed 12 clean inactive worktrees through `git worktree remove`; all
  detached heads were already ancestors of `main`, branch-backed worktrees kept
  their branches, and the two dirty worktrees plus official benchmark evidence
  were preserved. A development-root scan found no remaining LFS object store.
  Rounded Data-volume usage fell from 280 GiB to 262 GiB. Active applications,
  Codex/Claude sessions, Downloads, benchmark evidence and its official VM
  disks, project sources, and `.scratch/` were not deleted.
- 2026-08-21: PR #275 final head `52ddcba6` passed 19 remote checks with one
  intentionally skipped matrix placeholder and merged to `main` as
  `b62b78d704ec54b3ff43e333982eef713a33c8df`. CodeQL, unit-linux, full graph,
  clean-host, similarity, browser security, packaging, and required aggregation
  were green. A detached checkout of the exact merge had neither Git LFS nor
  local graph archives; stdlib hydration downloaded the 110,283,462-byte
  runtime and 295,151,086-byte full assets with exact manifest digests, then
  deep validation passed 79,958 nodes and 1,778,069 edges.
- 2026-08-21: Completed safe local cleanup without rewriting history or touching
  `.scratch/`. Removed all seven release-backed LFS cache objects
  (1,968,180,750 bytes), the now-untracked working graph pair (405,434,548
  bytes), and four stale Git temporary packs (197,656,624 bytes): at least
  2,571,271,922 bytes (2.395 GiB) freed. Removed obsolete checkout-local LFS
  configuration and empty cache metadata. No LFS-tracked path, `.git/lfs`
  directory, LFS config, or Git garbage remains; `git fsck --no-dangling`
  passed. The safe GitHub REST disable attempt returned HTTP 404 because that
  endpoint is not available for this personal repository. Remote object purge
  is now a prepared GitHub Support handoff, not a code or local-disk blocker.
- 2026-08-21: Began the safe Git LFS retirement on branch
  `codex/remove-git-lfs` from `a62a7ff6`. Independent audit counted 45 whole
  compressed archive objects totaling 13,025,281,486 bytes (12.131 GiB): 31
  full graph revisions and 14 runtime revisions. Twenty-two objects are already
  preserved as release assets; the current v1.0.21 pair (405,434,548 bytes)
  matches public release sizes/digests and verified attestations. The remaining
  historical snapshots are intentionally disposable because they are old tag,
  transient branch, or intermediate generated states and are not consumed by
  the published packages. Implemented a pinned five-asset release manifest and
  one stdlib resolver, migrated CI/publish/Hugging Face/local consumers and A/B
  proof preparation, removed tracking rules/four generated hooks/obsolete
  guard, and uninstalled the checkout-local LFS filters. Focused integrated
  suites report 326 and 113 passed. No history rewrite, remote purge, or local
  cache deletion has occurred; `.scratch/` and `.gate/worktrees` remain
  untouched.
- 2026-08-21: Completed aggregate verification and independent final audit for
  the LFS retirement tree. Repository-wide non-integration passed 8,812 tests
  with 5 skips. The first authoritative preflight attempt encountered the
  previously documented parallel-only deterministic-bridge request-count
  miss; the exact test passed three isolated reruns, and the full retry passed
  all 21 lanes, including clean-host install, strict docs, telemetry, release
  hydration, deep graph validation, similarity, browser security,
  reproducible wheel/sdist, and Twine. Independent audit first rejected two
  selective-hydration/source-contract defects, then accepted their repairs
  after adversarial refreeze with no P0-P2 findings. No local cache or remote
  LFS object has been deleted yet.
- 2026-08-21: Committed the accepted migration as `9e0fb368` and ran the
  committed-head gate in isolated clean worktrees. Eleven lanes progressed or
  passed, while the unit lane failed 123 benchmark/holdout fixtures because
  the runtime archive was correctly absent. This exposed a real difference
  between the previously hydrated development tree and clean CI checkouts.
  Added failing-first contracts, then made the local unit lane and both GitHub
  full-test jobs hydrate only `graph/wiki-graph-runtime.tar.gz` through the
  exact manifest before pytest. The three new red contracts and 181 affected
  CI/workflow tests are green; clean committed-head rerun is pending.
- 2026-08-21: Amended the clean-checkout repair into implementation commit
  `72d1ad73`. Its isolated unit lane selectively hydrated only the runtime
  archive and passed 8,806 tests with 5 skips and 92.16% coverage. To avoid
  repeating that identical five-minute lane, ran the other 11 local-fast lanes
  separately against the same exact commit; every lane passed. Together the
  two invocations cover the complete 12-lane committed-history gate. PR push,
  required remote checks, and merge are next.
- 2026-08-21: Opened PR #275. Fifteen specialized Tests checks passed while
  the long unit/graph jobs continued, but GitHub security aggregation opened
  high-severity CodeQL alert #23: `write_manifest` made the atomic temporary
  manifest world-readable with mode `0644`. Captured a failing-first permission
  regression, changed the manifest mode to owner-only `0600`, and passed 74
  focused tests plus Ruff, format, and mypy. Independent reviewer directly
  probed replacement and injected failure behavior and accepted with no P0-P2;
  the repair has not yet been pushed.
- 2026-08-16: Released CTX Fit 1.0.21. PR #271 merged as exact commit
  `38a33f8784e2bf408430a98fed81206c2cf39d00`; canonical Tests run
  `31914958343`, CodeQL run `31914958371`, and Hugging Face sync
  `31914958347` succeeded. The recovered annotated tag object
  `a7b8e78559fda1d44dca844393458272071ae89b` peels to that commit without a
  force update. Production publish run `31915534546` completed build, package
  provenance, graph provenance, CycloneDX attestation, release assets, and PyPI
  Trusted Publishing. Public release:
  `https://github.com/stevesolun/ctx/releases/tag/v1.0.21`; public package:
  `https://pypi.org/project/claude-ctx/1.0.21/`.
- 2026-08-16: Independently downloaded and verified the public artifacts. PyPI
  wheel SHA-256 is
  `aa0c6bd75412788df31f38fe43e3c6f5f6504a4e6a3b95d8126a3ddab9c5610a`;
  sdist SHA-256 is
  `fde373139a5de2f5543dcdc605b2533797a1c9c4fabfae357118bd778a5ed8f3`;
  release SBOM SHA-256 is
  `e83d0f62c6bc5207a6531a1bd2eaff485efd9babca214fa937cbb44cfaf6264d`.
  The SBOM is CycloneDX 1.6 for `claude-ctx` 1.0.21 with exact PyPI PURL and
  deterministic serial `urn:uuid:15633df7-79c8-5a2c-b5d9-bbebf39ac64e`.
  Sigstore/Rekor verification bound package provenance and the SBOM predicate
  to the exact tag, source commit, and pinned publish workflow. GitHub
  attestations are package provenance `40951384` / Rekor `2481664308`, graph
  provenance `40951389` / Rekor `2481664588`, and package CycloneDX binding
  `40951393` / Rekor `2481664719`. PyPI publish transparency indices are
  `2481670714` for the wheel and `2481670637` for the sdist. Versioned and
  unversioned PyPI JSON, Simple, and `pip index` all resolve 1.0.21. A fresh
  public install selected 1.0.21; `ctx --version`, `pip check`, and a minimal
  repository's read-only `ctx fit --json` profile passed.
- 2026-08-16: Archived the failed publish recovery evidence outside the
  repository at
  `/Users/steves/Steves_Files/Work/Research_and_Papers/ctx-release-recovery-31913053842`;
  its `SHA256SUMS` digest is
  `10cc282df0122fce3f99b6dff395d1b7764d0239c84adc915d3613751caaa5ed`.
  After offline attestation verification, deleted exact failed-run Actions
  artifacts `9254254768`, `9254254867`, and `9254257587`, plus exact GitHub
  attestation records `40948323` and `40948329`. Rekor entries `2480992277`
  and `2480992911` intentionally remain as transparency history. This freed
  about 420 MB of Actions storage, not Git LFS storage.
- 2026-08-16: Audited Git LFS. Only the full and runtime graph tarballs are LFS
  tracked; reachable history contains 45 unique objects totaling 12.131 GiB.
  Standard verified pruning has zero safe local candidates. The current pair is
  preserved as exact release assets, but 23 historical objects totaling about
  6.653 GiB lack release backup. GitHub continues billing remote LFS objects
  after pointer deletion/history rewriting, so safe quota recovery requires a
  reviewed release-asset migration followed by a GitHub Support purge; no
  destructive history rewrite was performed.

- 2026-08-16: Committed the independently accepted CycloneDX serial repair as
  `b3145c49`. Its committed fast gate passed 8,783 tests, 5 skips, 92.16%
  coverage, static checks, and clean-host/package contracts. The authoritative
  PR preflight then passed all 19 lanes, including the same 8,783-test unit
  equivalent, strict documentation, telemetry, similarity, browser security,
  clean-host installation, reproducible distribution build, and Twine.
  Reproducible artifact SHA-256 values are
  `40c8e88adb6ffcd6159ed30eddc25162db6e5e17bb6f0ac035ad2162857b70d7`
  for the wheel and
  `ce6fb0da869d8b3b5c6ff27327940f494b98f4c4efb34d466704983144368a6f`
  for the sdist. Only the user-owned `.scratch/` remains untracked; push and
  remote PR review are next.

- 2026-08-16: Merged PR #270 as exact main `2714f9be`; exact-main Tests run
  `31912497654` and CodeQL run `31912497612` succeeded. Explicitly deleted the
  still-unpublished old `v1.0.21` tag and created annotated object `bcd5632d`
  on that exact tested commit, without force-updating it. Publish run
  `31913053842` passed provenance, graph, stats, static/clean-host/canaries,
  reproducible build, distribution checks, wheel and all-runtime-extras smoke,
  deterministic PURL normalization, strict dependency closure, and artifact
  upload. Package/graph provenance attestations `40948323` and `40948329` were
  created, then the SBOM attestation failed closed: the valid CycloneDX 1.6
  file omitted optional `serialNumber`, while pinned `actions/attest` requires
  it. Release-assets and PyPI jobs skipped; GitHub release remains absent and
  PyPI 1.0.21 remains 404. The exact uploaded SBOM independently passes the
  official 1.6 schema. A failing-first repair now derives a content-addressed
  UUIDv5 serial, refuses conflicts, requires exact recomputation, and passes
  28 focused supply-chain tests plus the pinned action's three-field probe.
  Public inventory is now 8,780.

- 2026-08-16: Committed the accepted SBOM repair as `4ff0e890`. Its committed
  fast gate passed 8,779 tests, 5 skips, 92.16% coverage, static checks, and the
  clean-host contract. The authoritative PR preflight passed all 19 lanes:
  inventory/policy/static checks, the same 8,779-test unit equivalent, A-Z and
  compatibility canaries, clean-host installation, public docs contracts,
  strict MkDocs, telemetry, similarity, browser security, reproducible build,
  and Twine. Reproducible artifact digests were
  `748a0862e379a2dc0e8b8a671484225be588d80d9372e9412368377ac6cbfc32`
  for the wheel and
  `5c08024e5c88161ac7ed56f05edb4b7a26a91fc2861cc0188f0c98b96dac7a7e`
  for the sdist. Independent review accepted exact commit `4ff0e890` with no
  P0-P2 finding. Only the user-owned `.scratch/` remains untracked.

- 2026-08-16: Merged PR #269 as exact main `c1d8405b`; main Tests run
  `31910183189`, CodeQL, docs, and Hugging Face sync all succeeded. Created
  annotated `v1.0.21` tag object `2514da44` on that commit. Production publish
  run `31910688624` passed provenance, graph, stats, static, clean-host,
  canaries, reproducible build/package checks, wheel smoke, and all-runtime-
  extras smoke, then failed closed at SBOM validation: cyclonedx-bom 7.3.0
  omits the release root PURL. Zero workflow artifacts were
  uploaded; attest/release-assets/PyPI jobs skipped; GitHub release is absent;
  PyPI 1.0.21 remains 404. A failing-first generator-shaped regression now
  drives deterministic binding of the root and optional matching installed
  component without changing graph references; conflicting identities and
  duplicates fail closed. A fresh wheel-only all-runtime-extras environment
  reproduced the production root-only shape: both pinned-generator outputs
  passed strict closure validation after binding and remained byte-identical.
  The integrated release/workflow selector passed 190 tests, and the generated
  public inventory is now 8,776.

- 2026-08-16: Committed the deterministic inventory repair as `ca2b5799`.
  Its committed fast gate passed all lanes, including 8,775 tests, 5 skips,
  92.16% coverage, static checks, clean-host, and package contracts. The clean
  PR preflight passed all 16 lanes; reproducible wheel SHA-256 is `07bbd80a...`,
  reproducible sdist SHA-256 is `5f224abd...`, and Twine passed. Independent
  refreeze repeated 206 release/docs/workflow tests and proved the normalized
  8,772 inventory in both full-extra and fresh clean `[dev]` environments, with
  no P0-P2 findings. Only remote review/merge and new exact-main evidence remain
  before tagging.

- 2026-08-16: PR #267 merged as exact main `7cf8fb62`. Its main-push Tests run
  `31908356448`, CodeQL, Linux/macOS matrices, package/build lanes, and required
  zero-spend Ubuntu sandbox job all succeeded. Published and independently
  re-downloaded the validated graph cache prerelease
  `graph-artifacts-v1.0.21-7cf8fb62`; full/runtime/catalog assets match their
  committed size and SHA-256 contracts. Hugging Face retry hydrated from that
  cache, then exposed a release-stat environment dependency: 8,795 tests with
  optional extras versus 8,771 under clean `[dev]`. A failing-first regression
  now drives AST-based module-level `importorskip` normalization; focused tests
  and static checks pass, and the added regression makes the invariant public
  count 8,772. No `v1.0.21` tag, GitHub release, or PyPI version exists yet.

- 2026-08-15: Commit `c4aaa22e` passed all 11 committed fast lanes and all 19
  clean detached PR-preflight lanes, including 8,774 tests, 5 skips, 92.16%
  coverage, reproducible wheel `e9770d99...`, reproducible sdist `a0d86319...`,
  and Twine. Required Ubuntu job `94903820786` then started every Bubblewrap
  child and passed 8 of 10 adversarial checks. The two failures exposed an
  evidence-model distinction: writes beside the workspace and POSIX shm were
  private to Bubblewrap's new tmpfs/dev mounts, not host mutations. The active
  repair enforces the stronger host-backed single-writable-tree policy with a
  non-recursive root remount and proves Linux shm isolation by using the same
  name inside and outside the sandbox; macOS denial remains unchanged. Focused
  behavior/static checks and independent review are green; commit, gates, and
  exact-SHA Ubuntu remain.

- 2026-08-15: Required Ubuntu run `31840406705` reached the real Bubblewrap
  boundary and failed with AppArmor denying loopback setup. Three positive
  checks failed; seven negative tests had not proved their child started. The
  accepted repair loads the packaged capability-stripping
  `bwrap-userns-restrict` profile without disabling the global restriction,
  adds an exact child-start sentinel to isolation tests, and makes doctor/live
  setup execute a bounded no-network canary before any campaign. Public Ubuntu
  instructions preserve existing administrator policy and disclose that the
  profile applies to all `/usr/bin/bwrap` callers. Coordinator evidence is 64
  focused and 513 full Fit tests; independent refreeze repeated 513 Fit, 99
  focused, and 75 adjacent tests plus static/docs checks with no P0-P2. Exact
  committed gates and remote Ubuntu evidence remain.

- 2026-08-14: Committed Linux/provider/CI remediation as `10e47d37`. Its
  committed-history fast gate passed all 11 lanes in 323.45 seconds; the clean
  detached PR preflight passed all 19 lanes, including 8,769 tests, 5 skips,
  92.16% coverage, reproducible wheel `03cdaac7...`, reproducible sdist
  `1b488e9e...`, and Twine. Recording that result creates this state-only
  follow-up; proportionate exact-SHA checks and remote required CI remain.

- 2026-08-14: Repaired all eight `unit-linux` failures without weakening
  production boundaries. Semantic live-runner tests now inject their executor;
  the default model's provider and official release-verified price resolve
  without optional LiteLLM; missing `[harness]` refuses before trial setup; and
  a new required zero-spend Ubuntu lane installs `[harness]` plus Bubblewrap,
  runs ten real adversarial sandbox checks, and constructs without invoking the
  live driver. Full Fit passed 508 tests, workflow/CI contracts passed 137, and
  independent integration review accepted with no P0-P2. Generated inventory
  is current. Release remains stopped for commit, committed gates, and remote
  Ubuntu evidence.

- 2026-08-14: Security commit `2cc8667d` passed all 11 committed fast lanes and
  all 19 clean PR-preflight lanes; both remote CodeQL checks and 15 specialized
  PR checks passed. Remote `unit-linux` then failed eight environment-contract
  cases: two reached a missing Bubblewrap boundary and six expected optional
  LiteLLM harness metadata/pricing absent from `[dev]`. The required CI
  aggregate consequently failed. Dispatched non-overlapping Linux-sandbox and
  provider/pricing expert repair lanes; release remains stopped.

- 2026-08-14: Committed release candidate `0264cede` passed all 11 committed
  fast-gate lanes and all 19 clean PR-preflight lanes, including 8,761 tests,
  reproducible package construction, Twine, clean-host, docs, and static checks.
  Opened community PR #267. Remote CodeQL then found one high-severity
  world-readable applied-manifest permission issue. Captured a failing-first
  regression and repaired new manifest creation from `0644` to owner-only
  `0600`; 64 apply tests, 523 Fit/surface checks, docs/stat checks, and focused
  static checks pass. An independent reviewer accepted create/modify/rollback
  behavior with no P0-P2 findings. Release is stopped pending a new commit,
  repeated committed gates, and green remote CodeQL.

- 2026-08-13: Created this live state file after confirming that root
  `STATE.md` is intentionally frozen history. Recorded the merged base,
  independent release verdict, active parallel lanes, blockers, evidence
  freshness, required loops, and release stop condition.
- 2026-08-13: Reconciled the checkpoint with the shared worktree and active
  agents. Rejected the production driver's completion claim after a direct
  `codex sandbox :workspace` probe wrote to a sibling temporary directory.
  Added a stricter shared process boundary and recorded 6 focused security and
  audit regressions plus compile/Ruff evidence; provider integration and
  independent review remain open.
- 2026-08-13: Replaced the escaped production-driver wrapper with the shared
  host boundary. Direct adversarial probes now deny sibling writes and ambient
  temporary-secret reads; provider and environment-reuse focused suites are
  green. Separated dependency-setup network authority from network-disabled
  verification. Received spend/fairness and five-language task lanes for
  independent integration review; exact apply materialization remains active.
- 2026-08-13: Independent apply review rejected the first exact-material draft
  despite 63 green focused tests. Added red repair requirements for ambiguous
  marker refusal, stale-preview compare-and-swap, byte-preserving newlines,
  canonical material identity, complete treatment manifests, provider use of
  those exact bytes, and an actual run-time consumer of the applied sidecar.
- 2026-08-13: Independent spend/fairness review rejected the first repair.
  Began sealing authorization to the canonical executable plan, restricting
  simulation to the built-in runner, refusing invalid/over-cap costs, making
  JSON plan-only, expanding the human preview, and requiring an exact
  baseline-versus-challenger field with every declared trial slot complete.
- 2026-08-13: Reconciled this checkpoint against the live branch and agent
  roster. Recorded the completed five-language task lane, the active strict
  applied-configuration consumer, and the independent rejection of AGENTS
  mutation because it changes the winner's instruction preimage. The apply
  repair is now sidecar-only; activation and integrated exact-byte evidence are
  still open. No tag, package publication, or paid provider evaluation has
  occurred.
- 2026-08-13: Integrated the accepted sidecar-only apply result and strict
  ordinary-run activation. Added an exact applied baseline, a controlled-trial
  suppression flag to prevent an existing sidecar contaminating every Fit arm,
  and one canonical renderer for trial and applied instruction/capability
  bytes. A 282-test integration selector was green before later spend edits.
- 2026-08-13: Reopened spend/fairness after independent adversarial refreeze
  found unsigned reliability, duplicate identity, exact-cap, unknown-cost,
  precision, and simulator-subclass defects. Bound floor/task/report structure
  into execution and recommendation, removed report-digest relabelling, and
  started reconciling strict fixtures. Current selector: 83 passed / 15 failed.
  Two independent reviewers are now active in parallel; no full gate, tag,
  publication, or paid provider call has been attempted.
- 2026-08-13: Reconciled strict comparison and sidecar-only CLI fixtures and
  closed the reviewer repros for exact caps, unknown/invalid/over-cap cost,
  plan/task/floor binding, bool slot identity, simulation identity, and
  full-precision selection. Focused spend selector is now 100 passed with
  Ruff/format/mypy/diff checks green; independent final refreeze is active.
- 2026-08-13: Independent activation/security review found a new P0: editable
  source could terminate the repository verifier successfully before tests
  completed and receive `verified`. A separate writer now owns only the live
  runner and its focused tests; the coordinator continues release integration
  in parallel. Release remains stopped.
- 2026-08-13: Spend/fairness received final independent ACCEPT after 106 focused
  tests and direct adversarial probes; no findings remain in that lane. The
  integrated Fit suite is now 440 passed. The verifier completion-witness
  repair is 33 tests/static green and under independent refreeze; non-Python
  verification currently fails closed pending an honest adapter design.
- 2026-08-14: Completed a tracker-repaired repository-wide checkpoint with
  8,707 passed and 5 skipped, then correctly marked it stale when independent
  refreeze found two new P0s: the first-use baseline omitted installed
  capabilities, and the sandbox exposed ambient host reads/local sockets. A
  failing-first baseline regression now passes with exact simple repository
  skill material; unreproducible current agents/MCP/tool layouts abstain. The
  applied-model P1 repair is focused-green, while sandbox and multi-language
  dependency-scope writers continue in parallel. Release remains stopped.
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 472kUpdated Claude CodeCodexVerifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 322kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 322kUpdated Claude CodeCodexVerifiedInstall
executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is availableStars: 322kUpdated Claude CodeCodexVerifiedInstall