admin/claw
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: add generated scene skill platform hardening

Browse Source
This commit is contained in:
木炎
2026-04-21 23:19:06 +08:00
parent 118fc77935
commit 956f0c2b68
439 changed files with 61974 additions and 3645 deletions
Download Patch File Download Diff File Expand all files Collapse all files

130
docs/superpowers/specs/2026-04-19-structured-fail-closed-improvement-roadmap-design.md Normal file
View File

@@ -0,0 +1,130 @@
# Structured Fail-Closed Improvement Roadmap Design
> Date: 2026-04-19
> Status: Draft
> Upstream Reconciliation: `tests/fixtures/generated_scene/full_sweep_status_reconciliation_2026-04-19.json`
## Intent
Turn the `48` structured fail-closed records from the reconciled `102` sweep into a governed improvement roadmap.
The objective is not to weaken gates or inflate `auto-pass`. The objective is to classify contract gaps, identify the highest-value bounded correction slices, and then improve generic scene-to-skill conversion where evidence can be recovered safely.
## Current Reconciled Baseline
After status reconciliation, the `102` scene set is:
| Reconciled status | Count |
| --- | ---: |
| `auto-pass` | 48 |
| `fail-closed-known` | 48 |
| `adjudicated-valid-host-bridge` | 4 |
| `source-unreadable` | 2 |
| `missing-source` | 0 |
| `unsupported-family` | 0 |
The `4` raw route conflicts are no longer unresolved route bugs. They are valid host-bridge workflows.
This roadmap therefore focuses on `fail-closed-known = 48`.
## Fail-Closed Buckets
| Inferred archetype | Reason | Count |
| --- | --- | ---: |
| `paginated_enrichment` | `workflow evidence is incomplete before package generation` | 35 |
| `local_doc_pipeline` | `workflow evidence is incomplete before package generation` | 5 |
| `multi_mode_request` | `workflow evidence is incomplete before package generation` | 4 |
| `single_request_enrichment` | `workflow evidence is incomplete before package generation` | 2 |
| `host_bridge_workflow` | `workflow evidence is incomplete before package generation` | 1 |
| `page_state_eval` | `bootstrap_target` | 1 |
The first priority is the `35` `paginated_enrichment` records because they are the largest bucket and map to the most important currently generic workflow family.
## Scope Guardrails
In scope:
1. classify the `48` structured fail-closed records by missing contract piece
2. prioritize bounded correction slices
3. implement bounded evidence recovery only after classification shows repeated recoverable patterns
4. keep all fail-closed semantics intact
5. rerun a bounded follow-up sweep after corrections
Out of scope:
1. adding new scene families
2. starting `G4/G5`
3. login recovery
4. full browser host runtime transport
5. local document attachment runtime
6. auto-promoting scenes into the execution board
7. weakening readiness gates to increase pass counts
8. reopening the already adjudicated `4` valid-host-bridge workflows
9. handling the `2` remaining timeout records in this roadmap
## Workstreams
1. `WS1` Fail-Closed Inventory and Gap Taxonomy
2. `WS2` G3 Paginated Enrichment Contract Recovery
3. `WS3` Small-Bucket Contract Recovery
4. `WS4` Bootstrap Target Isolation
5. `WS5` Follow-Up Sweep and Coverage Delta
## Gap Taxonomy
Every structured fail-closed record must receive one primary missing-contract label:
1. `main_request_missing`
2. `pagination_plan_missing`
3. `enrichment_request_missing`
4. `join_key_missing`
5. `export_plan_missing`
6. `mode_matrix_missing`
7. `mode_request_contract_missing`
8. `single_request_enrichment_contract_missing`
9. `host_bridge_contract_missing`
10. `local_doc_contract_missing`
11. `bootstrap_target_unresolved`
12. `mixed_or_ambiguous_contract_gap`
Secondary labels may be added, but every record must have exactly one primary label.
## Correction Strategy
Corrections must be pattern-based, not scene-by-scene.
Allowed correction types:
1. bounded evidence extraction for repeated field names or workflow structures
2. bounded IR fallback only when evidence is explicit and traceable
3. more specific fail-closed reason reporting
4. regression tests for each recovered pattern
Forbidden correction types:
1. hard-coding a scene name to pass
2. converting fail-closed records to pass without closing the contract
3. broad route-precedence rewrites
4. disabling or relaxing gates
## Expected Outputs
1. `tests/fixtures/generated_scene/structured_fail_closed_inventory_2026-04-19.json`
2. `tests/fixtures/generated_scene/structured_fail_closed_improvement_followup_2026-04-19.json`
3. `docs/superpowers/reports/2026-04-19-structured-fail-closed-inventory-report.md`
4. `docs/superpowers/reports/2026-04-19-structured-fail-closed-improvement-coverage-delta-report.md`
5. `docs/superpowers/reports/2026-04-19-structured-fail-closed-improvement-roadmap-closure-report.md`
## Acceptance Criteria
1. all `48` fail-closed records are inventoried
2. all `48` records have exactly one primary missing-contract label
3. the `35` `paginated_enrichment` records are split into actionable G3 gap groups
4. implementation, if performed, is limited to repeated recoverable patterns
5. no adjudicated host-bridge record is reopened
6. follow-up results are measured against the reconciled baseline
7. execution board status remains unchanged
## Completion Signal
The roadmap is complete when the `48` structured fail-closed records are no longer a single broad bucket and the follow-up sweep quantifies whether bounded evidence recovery improved safe conversion coverage.