Fix skills lifecycle gates and retain skills qualification evidence - #719

Scope

A worker reuse entry could install a centrally held skill because it inherited listing gates but kept its own trial status. Resolve central held/pruned status through the shared installer, make OpenHands use the same eligible set, and record both manifest hashes.

Map installed skills across all 13 lifecycle tasks, publish the degraded sweep as STOPPED, and retain the assigned upstream/native readiness results. Research-skill round 2 records BY_DESIGN for new vendor candidates: two blind Sol decisions, Astra and a final completeness critic converge on configuring the existing owner, with no material split or conditional Opus verification.

Register the already-installed native-stack-research entry at an immutable lifecycle directory pin, preserving the landed embedded-DeerFlow source byte for byte (1,602 bytes, SHA256 313f41cbbb7ba5cbe9733d7b9c4caf85bd4fabf4c7a4cb5b40731f5449b941eb). Set its manifest listing on and Codex enabled, retain Unknown audits, align repository template metadata, and explicitly exclude the host-specific entry from worker copying pending a qualified bridge. This repairs ownership; it does not certify useful research or model visibility. No new vendor candidate, shared-host install, plan apply or live client setting change.

Publish follow-up 1 wf_c2f7dbc8-129 through the unchanged native harness: two tasks, 29 proposals, 10 vote survivors, 8 bound reviews and 2 CI/PR source stops. Retain the stopped record and discovery timeout; no new candidate is selected. Correct the existing research skill's catalog source/task mapping, preserving the retained sweep's frozen inputs. Source reviews and local checks establish provenance/consistency, not behavioral or runtime acceptance.

Publish follow-up 2 wf_4b9c7d98-ffd as a stopped owner publication: CI/PR, agent-docs and review; 41 proposals, 17 retained vote leads, 14 bound source reviews and three source stops. Two judged SKILL.md hashes are null; openai/plugins@verify-fix is absent from the CLI's default discovery locations at its valid pin without --full-depth. Native source review exits 1 and completed-result construction exits 2; the supported stopped record retains all leads and reopens affected source gates. CI/PR also retains a GPT discovery timeout. Converted artifacts preserve actual bytes; one sanitized discovery note and producer-reported initial convert exit 3 remain explicit. Requested GPT model/effort are metadata observations; OmniRoute transport is producer-reported. No candidate is selected, benchmarked or installed by this publication.

• Base: rebased onto fixed 424b8a778222ae8857aade42c49ae6f7a13b3138. Original full-sweep inputs retain their ec0b8fd8 basis; FU1 baselines retain e7d94e7b; FU2 baselines use fixed 424b8a77. Main's 236 anti-pattern table rows remain byte-identical and in order, followed by this lane's three rows. Main's client-budget prose and template skillListingBudgetFraction: 0.05 are preserved. Registry rebuilt from fixed main, unowned rows and other keys preserved; shared ledger and registry committed last.
• Lane: lane:foundation.
• Owned paths: shared skills installer, OpenHands resolver and tests, central skills manifest and unchanged package, matching template metadata, worker manifest/documentation, foundation decisions and retained evidence.
• North-star action: useful public research and skills-guided complex-system work across native clients and research workers.
• Keep draft until the command center's Claude cross-family read. The command center owns landing; readiness-runner is the sole shared-host plan-apply writer.

SOTA sources

• FU2 and the compatible rebase follow native-agent-stack@424b8a77, owner publication steps 9–12, judged-source stops, native stopped-record recipe, registry hot-file protocol, and evidence classes. All 17 source dispositions remain in the FU2 source-review index; only 14 bind the stopped result. The retained completeness critic's missed baselines and modality leads feed the next task-scoped sweep, without asserting those leads are qualified.

• FU1 follows native-agent-stack@e7d94e7b, owner publication steps 9–12, judged-source stops and native stopped-record recipe. Catalog mapping follows existing native completeness invariants using the already published 05e3ead research package.

• Supported installation remains vercel-labs/skills@7407f3893ad4dceab546ac002c3ef806e4000c73, README (Skills 1.7.0), through the repository lifecycle; no replacement installer.

• Shared installation/discovery follows Agent Skills client integration and Python's direct-file importlib recipe. OpenHands keeps SDK v1.50.1's native skill loader.

• Native publication follows native-agent-stack@ec0b8fd8, landscape-sweep README, steps 9–12, judged-byte binding, make_result.py:160, completed-result refusal:244, manual stopped record, and ledger review binding.

• Visibility sources remain Codex rust-v0.160.0 render.rs:126, prompt_debug.rs:88, skills_config.rs:20, Claude Code v2.1.289 and official description-budget diagnostics. These support dated readiness observations, not a new model run.

• Worker provisioning sources include DeerFlow v2.1.0 storage, Inspect 0.3.273 skills example and Harbor v0.23.0 job skills.

• Candidate pins, judged SKILL.md hashes and license metadata are retained in the converted artifacts and 37 source reviews. Provenance is not license approval or supporting-code acceptance. Future qualification uses anthropics/skills@8a1541c4a3ffa5a20a5a91de0dcf3f0bab1d1ef4, skill-creator paired workflow, then the repository lifecycle and fresh client/worker discovery. No custom benchmark runner.

• Native readiness uses Trail of Bits' unchanged validation at 82fe8226, including Claude loadability:149 and Codex:161; Skills CLI's CI at 7407f389; and engineering package scripts at d81f3a18.

• The fictional authoring probe reuses the existing Anthropic creator above and the byte-identical official Codex rust-v0.160.0 initializer:175 and validator:208. Codex transport follows the repository's documented keyless-loopback placeholder; no credential or persistent client configuration is changed.

• Existing research source is frozen at native-agent-stack@c148e049:research-harnesses-skill.md:1; package at 05e3ead0:adoption/skills/native-stack-research, tree 2d233cad78ad1d2da70721fd46fbe20088909558. No draft Repair native install-plan acceptance after the hcom relaxation #723 bytes.

• Research round 2 follows repository-quality rule@c148e049:17, reuses input pack@cb55ae1c, and uses native Codex 0.160.0, with the reviewed codex_job.py@46c00618:686 dispatch contract.

• Worker research exclusion follows the existing central reuse/exclusion contract@c148e049 and frozen host-script requirements@05e3ead0:8. Catalog metadata changes follow the existing listing decision@c148e049.

Evidence-class table

Readiness artifacts: evidence/artifacts/skills-readiness-20261005/, including local-checks.json with exact returned-output hashes and retained failed checks.
Raw conversation streams stay in owned private scratch; bounded results, exact raw hashes, native counters and declared sanitizations are retained. Native tests, our integration checks and fictional fixtures remain distinct.

Local commands run

Current FU2 checks on fixed main 424b8a77 (local integration; 1,162 tests, ten skips):

python3 -m unittest tests.test_saturation_ledger tests.test_landscape_sweep_harness tests.test_landscape_sweep_skills
# exit 0; 498 tests, 6 skips
python3 -m unittest tests.test_skills_manifest tests.test_install_skills tests.test_skills_status tests.test_render_config tests.test_adoption_docs_consistency tests.test_runtime_worker_skills
# exit 0; 272 tests, 1 skip
python3 -m unittest tests.test_runtime_worker_openhands tests.test_runtime_worker_openhands_resolver tests.test_runtime_worker_openhands_push_gate tests.test_codex_worker_skill
# exit 0; 392 tests, 3 skips
python3 scripts/saturation_ledger.py --append evidence/artifacts/skills-lifecycle-landscape-20261005-fu2/RESULT-input.json
# exit 0; stopped result retained, ledger immediately re-registered
python3 scripts/saturation_ledger.py --check --base 424b8a778222ae8857aade42c49ae6f7a13b3138
# exit 0; eight records, chain and bindings verified
python3 scripts/validate.py
# exit 0; 69 components, 10427 hashed files, 4 profiles, 208 receipts
# integrity/scope checks only; no live provider or GPU execution

Native FU2 source_reviews.py exits 1 with 14 bound / three stopped; make_result.py exits 2. Both actual stderr files are retained. The stopped recipe appends without counting or resetting saturation. The one-note sanitization and producer-reported initial converter refusal remain in the provenance receipt. Final registry reconstruction registers 174 prior lane paths plus 30 FU2 paths through the official API, preserving unowned main rows and all other keys. Raw logs stay private; their digests and independent review are in checks.json.

Earlier retained unit checks:

Current FU1 checks:

python3 -m unittest tests.test_saturation_ledger tests.test_landscape_sweep_harness tests.test_landscape_sweep_skills
exit 0; 498 tests, 6 skips. Initial 2 failures retained; corrected existing research catalog source/task mapping.

python3 scripts/saturation_ledger.py --append evidence/artifacts/skills-lifecycle-landscape-20261005-fu1/RESULT-input.json
exit 0; status stopped. Initial unsupported trigger rejection retained; corrected to native retained_failure with exact source-stop refs.

python3 scripts/saturation_ledger.py --check --base e7d94e7bcc074edba093994c8853e59b650ce7e0
exit 0; 7 records, chain intact, bindings verified.

python3 scripts/validate.py
final exit 0; 69 components, 10,346 hashed files, 4 profiles, 204 receipts; integrity/scope only.
Initial stale appended-ledger hash failure retained; refreshed through official register_file.

git diff --check
git diff --cached --check
exit 0

Earlier registration commands below are retained historical checks at their original inputs:

python3 scripts/validate.py
exit 0; 69 components, 10,293 hashed files, 4 profiles, 204 receipts
Final current files verified after fixed-base registry rebuild.

python3 -m unittest tests.test_skills_manifest tests.test_install_skills tests.test_skills_status tests.test_render_config tests.test_adoption_docs_consistency tests.test_runtime_worker_skills
exit 0; 272 tests, 1 skip; rerun after the final rebase/pin.

python3 -m unittest tests.test_runtime_worker_openhands tests.test_runtime_worker_openhands_resolver tests.test_runtime_worker_openhands_push_gate tests.test_codex_worker_skill
exit 0; 392 tests, 1 skip; local/synthetic worker consumers.

python3 tools/adoption/install_skills.py --print-codex-config
exit 0; no rule disables native-stack-research; nonmutating rendering only.

git diff --check
exit 0

The unchanged pinned upstream quick_validate.py returns Skill is valid!, exit 0. Earlier assigned checks (457 skills/worker tests with three skips; 61 docs/release tests with one skip) retain their original inputs and evidence class. The initial new metadata suite failed four checks, corrected by the catalog count and explicit worker disposition. The registry attempt against a moving shared origin ref failed; rebuilt against fixed base095d4fad, with both attempts retained.

Upstream readiness commands and returned results are in the readiness receipts: unchanged Trail of Bits validate.yml jobs; [email protected] frozen install, test and type-check; native creator init/quick_validate/paired workflow/aggregate/viewer. Failed attempts remain retained. Historical sweep/ledger checks do not constitute new upstream acceptance. Release/re-pin work remains a command-center dependency.

Decision record

Current FU2: docs/decisions/2026-10-05-skills-landscape-sweep-fu2.md; comparison to overturn: judged bytes and pins resolved through the native source gate, then a frozen upstream skill-creator paired benchmark and cross-family review. No source workaround or benchmark is substituted for this stopped publication.

• docs/decisions/2026-10-05-skills-landscape-sweep-fu1.md

• docs/decisions/2026-10-05-research-skill-round2.md

• docs/decisions/2026-10-05-skills-lifecycle-readiness.md

• docs/decisions/2026-10-05-skills-landscape-sweep.md

• docs/decisions/2026-10-05-skills-native-readiness.md

The research entry stays at its landed embedded contract until #723 lands; any changed contract requires a normal lifecycle re-pin and changed-skill qualification. Three audits remain Unknown; the prospective skills.sh locator is not a returned audit. No new skill A/B, useful research acceptance or fresh model discovery is claimed by round 2. Fresh Claude/Codex discovery and worker bridge qualification wait for reviewed lifecycle installation by the sole host writer.

The original full sweep reopened twelve tasks. FU1 has scoped research voting/source closure for three leads; CI/PR retains its discovery timeout and two source-review stops. Co-op run 2 covers CI/PR, agent-docs and review. Remaining follow-up outputs stay with the co-op. Survivors remain leads until task-specific paired qualification and review. MinerU registration and Dagu stray-skill disposition wait on #719; no host install or client setting change follows this draft.

Host evidence

No new evidence/hosts/ receipt or platform-status promotion. Retained raw outputs and independent read-only reviews are linked in the artifact packages.

Checklist

• No changed GitHub Actions or new workflow permissions.
• No credential value/file read, printed or copied; presence-only helper used.
• No new paid hosting/subscription/billing surface.
• Peer worktrees and files preserved; only the owned lane checkout changed.
• One draft PR, one lane label; shared registry rebuilt from frozen main and committed last.