Dynamic Agent Skill Research

This repository separates the governed ChatGPT Project upload pack from the evolving research workflow.

Research dashboard

Open index.html locally to explore the evidence, matched-run results, runtime workflow, evaluation design, and structured improvement loop in a standalone interactive dashboard.

Current evidence shows that prompt-defined ephemeral agent composition is feasible on the observed Codex surface and usable as an explicit-only technical preview. DA-014 validated a routing-contract improvement at plan stage only. DA-022 scored 0 intention-to-treat from instrumentation failure, not model quality; DA-023 had no live run; and DA-024 closed checksums but failed semantic verification despite an actually unchanged workspace. DA-025 then passed its privacy-safe packet checks but observed no orchestration: only wait, no spawn, child, or linkage, and five unclassified errors. The permanent JSONL stop fired and G2 remains closed. Dynamic superiority and default adoption are not established.

Start here

• index.html — dependency-free interactive research dashboard.
• dynamic-agent-skill-research/ — validated 15-file ChatGPT Project initialization pack and canonical CSV ledgers.
• research-workflow/PROFESSIONAL_RESEARCH_CONTINUATION_PROMPT.md — execution specification for the current research epoch.
• research-workflow/PROFESSIONAL_RESEARCH_TO_USER_SKILL_PROMPT.md — professionalized execution contract for reaching a defensible conclusion and user Skill.
• research-workflow/PROFESSIONAL_RESEARCH_UNBLOCK_PROMPT.md — bounded authority, workflow, stop rules, and acceptance criteria for this unblock epoch.
• research-workflow/CONCLUSION-AND-REMAINING-WORK-2026-07-30.md — conclusion ladder, release lane, research lane, gates, and stop conditions.
• research-workflow/USER-SKILL-BETA-v0.1.md — exact-hash user preview, quickstart, validation, rollback, limits, and public-release blockers.
• research-workflow/release/RELEASE-MANIFEST.json — versioned Beta identity, exact file hashes, discovery policy, validation state, and release blockers.
• research-workflow/release/verify-user-skill-release.mjs — read-only project/user installation and release-integrity verifier.
• research-workflow/README.md — dependency graph, sequential/parallel paths, gates, and completion rules.
• research-workflow/EVIDENCE-MAP-2026-07-28.md — preliminary primary-source synthesis and unresolved claims.
• research-workflow/EVALUATION_PROTOCOL-v0.3.md — calibrated three-arm protocol with an explicit instrumentation gate.
• research-workflow/NEXT_ACTION-DA-001.md — completed runtime-feasibility probe definition.
• research-workflow/NEXT_ACTION-DA-002.md — recommended protocol calibration before any prototype.
• research-workflow/RESEARCH-UPDATE-2026-07-29.md — primary-source refresh, runtime findings, counterevidence, and current decision.
• research-workflow/DA-003-VALIDATION-2026-07-29.md — trace-adapter evidence matrix and remaining acceptance gaps.
• research-workflow/NEXT_ACTION-DA-003.md — immutable instrumentation checkpoints, the fired DA-025 permanent stop, and the optional new-boundary path.
• research-workflow/NEXT_ACTION-DA-006.md — preserve the stopped DA-011 epoch and govern a separately frozen compatibility batch.
• research-workflow/experiments/DA-006-source-isolation/ — independently reviewed 24-row plan-stage matrix, schemas, validator, and runner; DA-011 stopped after its first schema-rejected row.
• research-workflow/reviews/REVIEW-LOOPS-2026-07-29.md — finding-to-regression-to-repair evidence for DA-003 and DA-006.
• research-workflow/instrumentation/ — create-only saved-rollout adapter, fixtures, tests, and usage notes.
• research-workflow/calibration/ — six fixtures, answer keys, run schema, rubric, runbook, and DA-002 record.
• research-workflow/results/DA-001-2026-07-28/RESULT.md — observed two-agent and no-spawn probe result.
• candidate-skills/compose-task-agents/ — explicit-invocation ephemeral role compiler and orchestration technical preview; comparative performance is not yet established.
• research-workflow/results/DA-004-candidate-skill-2026-07-28/RESULT.md — candidate implementation and three-fixture forward-test record.
• research-workflow/results/DA-007-user-skill-forward-tests-2026-07-30/RESULT.md — fresh root, capacity-fallback, and prompt-injection forward tests.
• research-workflow/results/DA-008-clean-clone-skill-discovery-2026-07-30/RESULT.md — public clean-clone discovery, validation, and explicit invocation smoke test.
• research-workflow/results/DA-009-beta-hardening-2026-07-30/RESULT.md — fail-closed validator repair, independent review, refreshed hashes, and root/dynamic smokes.
• research-workflow/results/DA-010-DA003-F01-acceptance-2026-07-30/RESULT.md — six reconciled root-only canaries and the evidence for keeping G2 closed.
• research-workflow/results/DA-011-DA006-plan-stage-2026-07-30/SUMMARY.json — preserved one-row schema rejection, protocol stop, and 23 unattempted rows.
• research-workflow/results/DA-012-user-scope-install-2026-07-30/RESULT.md — reversible user installation and fresh out-of-repository CLI discovery smoke.
• research-workflow/results/DA-014-DA006-plan-stage-v3-scored-2026-07-30/ — 24-row routing-contract ablation; plan-stage evidence only.
• research-workflow/results/DA-018-DA003-F04-raw-identity-canary-2026-07-30/ — immutable pre-model feature-flag failure.
• research-workflow/results/DA-020-DA003-F04-integrity-gated-canary-2026-07-30/ — invalid response-schema stop before a valid delegated comparison.
• research-workflow/results/DA-022-DA003-F04-semantic-probe-canary-2026-07-30/ — B-only intention-to-treat score 0 from instrumentation failure; no C run or model-quality conclusion.
• research-workflow/reviews/DA023-PREFLIGHT-INDEPENDENT-REVIEW.json — rejected open-world authorization preflight; no live DA-023 run.
• research-workflow/results/DA-024-closed-auth-collab-trace-probe-2026-07-30/ — unscored negative instrumentation packet with checksum closure and semantic failure.
• research-workflow/reviews/DA024-POSTRUN-INDEPENDENT-REVIEW.json — immutable DA-024 HOLD and one diagnostic-only successor authorization.
• research-workflow/results/DA-025-privacy-safe-diagnostic-probe-2026-07-30/ — terminal unscored diagnostic; packet checks passed, orchestration was unobserved, and JSONL probing is retired.
• research-workflow/reviews/REVIEW-DA-010-2026-07-30.md — independent HOLD for DA-003/G2 after the bounded root-only canary pass.
• research-workflow/reviews/REVIEW-DA-011-2026-07-30.md — independent packet-integrity PASS and DA-006 restart HOLD with provenance limits.
• research-workflow/reviews/REVIEW-DA-012-RELEASE-2026-07-30.md — independent release-verifier PASS and public-release HOLD.
• .agents/skills/compose-task-agents — repository discovery entry for explicit $compose-task-agents invocation; implicit invocation is disabled.
• research-workflow/evaluation/ — frozen arm definitions, matched-demonstration runbook, and grader tooling.
• research-workflow/IMPROVEMENT_LOOP-v1.md — versioned single-change improvement, review, promotion, and rollback workflow.
• research-workflow/results/DA-005-F04-matched-demo-2026-07-28/ — complete three-arm demonstration artifacts.

Do not copy the entire repository into a ChatGPT Project. Follow dynamic-agent-skill-research/UPLOAD_MANIFEST.md and complete the pre-upload safety check first.