From e2ee923597785a7589e6542cbcaa161810c1bf34 Mon Sep 17 00:00:00 2001 From: wantsong Date: Sat, 20 Jun 2026 02:07:37 +0800 Subject: [PATCH] docs: define m2 prompt authority boundary --- PROJECTS.md | 24 ++- docs/DECISIONS.md | 43 +++++ docs/PROMPT_AUTHORITY_BOUNDARY.md | 124 ++++++++++++++ ...0-m2-cognitive-runtime-prompt-contracts.md | 156 ++++++++++++++++++ requirements/ccpe/README.md | 45 +++++ ...run-logger-and-adapter-manifest-checker.md | 121 ++++++++++++++ requirements/skills-vault/README.md | 43 +++++ 7 files changed, 547 insertions(+), 9 deletions(-) create mode 100644 docs/PROMPT_AUTHORITY_BOUNDARY.md create mode 100644 requirements/ccpe/2026-06-20-m2-cognitive-runtime-prompt-contracts.md create mode 100644 requirements/ccpe/README.md create mode 100644 requirements/skills-vault/2026-06-20-m2-runtime-run-logger-and-adapter-manifest-checker.md create mode 100644 requirements/skills-vault/README.md diff --git a/PROJECTS.md b/PROJECTS.md index 834323f..c181454 100644 --- a/PROJECTS.md +++ b/PROJECTS.md @@ -2,9 +2,9 @@ ## Active Phase -Current phase: M0-M1 startup. +Current phase: M2a boundary correction. -The active deliverable is a small file-first foundation for a cognitive-processing runtime. The repository should stay closer to runtime setup than governance evidence. +The active deliverable is the authority boundary for M2 prompt/runtime work. The repository should not create canonical prompts, canonical runtime specs, or Local CCRA review artifacts during M2a. ## Source Of Truth @@ -22,18 +22,24 @@ The old project is not an upstream branch and not a source tree to merge. 1. M0: project boundary freeze. 2. M1: model-management kernel. -3. M2: prompt/runtime skeleton. -4. M3: reader translation prompt and before/after translation check. -5. M4: first real sample run. -6. M5: first Local CCRA milestone review. +3. M2a: prompt authority boundary and supplier requests. +4. M2b: product-local runtime adapters and one minimal manual sample. +5. M3: reader translation refinement and before/after translation check. +6. M4: first full real sample run. +7. M5: first Local CCRA milestone review. -Do not skip from M1 into sample execution without an explicit owner instruction. +Do not skip from M2a into canonical prompt creation. If `prompts/` is created during M2b, files must be product-local draft adapters with explicit non-canonical frontmatter. ## Supplier Boundaries -This repository may record requests for external reusable tooling later, but M0-M1 should not create a supplier-request channel unless a real missing tool blocks progress. +This repository records supplier requests under: -Reusable policy, reviewers, and broader CCRA method should stay outside this product repo unless the owner explicitly asks to localize a milestone protocol. +- `requirements/ccpe/` for canonical Prompt Cards, Agent Specs, Runtime Specs, model-call protocols, evaluation rubrics, and Local CCRA review protocol. +- `requirements/skills-vault/` for deterministic reusable automation such as adapter validators, run loggers, manifest checkers, and review context builders. + +Reusable policy, reviewers, prompt authority, and broader CCRA method stay outside this product repo unless the owner explicitly asks to localize a product-only adapter. + +Do not copy canonical CCPE artifacts into this repository. Do not implement reusable automation locally when the correct action is a supplier request. ## Naming diff --git a/docs/DECISIONS.md b/docs/DECISIONS.md index 0cc7276..9ba2118 100644 --- a/docs/DECISIONS.md +++ b/docs/DECISIONS.md @@ -62,3 +62,46 @@ Reason: the system needs model identity and call boundaries before it needs auto Decision: reader translation is a first-class product layer, even though its prompt is not created in M0-M1. Reason: internal model output is not automatically understandable or usable by frontend readers. + +## D011: M2 Prompt Authority Boundary + +Decision: M2 must not directly implement GPT-authored prompt plans as canonical prompts in this repository. + +Reason: canonical Prompt Cards, Agent Specs, Runtime Specs, model-call protocols, evaluation rubrics, and Local CCRA review protocol belong in `ccpe-system`. + +## D012: Product-Local Adapter Scope + +Decision: this repository may later hold only product-local draft adapters, sample runs, product workflow glue, and output formats. + +Reason: product-local adapters are implementation-facing runtime surfaces, not canonical CCPE prompt authority. + +## D013: Mandatory Adapter Frontmatter + +Decision: if `prompts/` is created before CCPE supplies canonical contracts, every prompt-like file must declare: + +```yaml +authority: product_local_adapter +status: draft +canonical_source: pending_ccpe +not_canonical_ccpe_prompt: true +``` + +Reason: future sessions must not confuse product-local adapters with canonical CCPE artifacts. + +## D014: CCPE Request Channel For M2 Contracts + +Decision: M2 canonical prompt/spec/rubric/protocol needs are recorded under `requirements/ccpe/`. + +Reason: missing CCPE-owned authority should become a supplier request, not an ad hoc local implementation. + +## D015: Skills-Vault Request Channel For Deterministic Tools + +Decision: deterministic M2 tooling needs are recorded under `requirements/skills-vault/`. + +Reason: reusable automation such as adapter metadata validation, run logging, and manifest checking belongs in `skills-vault`, not in this product repository. + +## D016: Local CCRA Waits Until After A Minimal Manual Sample + +Decision: Local CCRA must not start during M2 boundary correction. + +Reason: Local CCRA should review milestone output after boundary correction and one minimal manual sample, not review the boundary correction itself as a new round. diff --git a/docs/PROMPT_AUTHORITY_BOUNDARY.md b/docs/PROMPT_AUTHORITY_BOUNDARY.md new file mode 100644 index 0000000..c1df88a --- /dev/null +++ b/docs/PROMPT_AUTHORITY_BOUNDARY.md @@ -0,0 +1,124 @@ +# Prompt Authority Boundary + +status: draft_m2_boundary_correction +date: 2026-06-20 + +## Purpose + +M2 must not directly implement the GPT-authored prompt plan as canonical prompts in this repository. + +The M2 correction is to separate product-local runtime work from CCPE-owned canonical prompt and runtime authority. + +## Authority Split + +| Surface | Owner | This repository may store | This repository must not store | +| --- | --- | --- | --- | +| Product runtime | `Cognitive-OS-Wantsong` | product-local adapters, manual run records, sample inputs/outputs, product workflow notes, product output formats | canonical CCPE prompt cards, Agent Specs, Runtime Specs, CCRA reviewer specs | +| Canonical prompt/runtime authority | `ccpe-system` | supplier requests under `requirements/ccpe/` | copied canonical artifacts treated as local truth | +| Deterministic automation | `skills-vault` | supplier requests under `requirements/skills-vault/` | reusable Skill source, generic validators, reusable CLIs | + +## Product Repository Scope + +This repository owns: + +- product-specific runtime adapters; +- sample run inputs and outputs; +- manual run records; +- product-specific workflow glue; +- local output formats for internal trace and reader-facing explanation; +- product-side requests to CCPE and skills-vault. + +This repository does not own: + +- canonical Prompt Cards; +- Agent Specs; +- Runtime Specs; +- model-call protocols; +- evaluation rubrics; +- Local CCRA canonical review protocol; +- reusable deterministic automation tools. + +## M2 Prompt Rule + +Do not create canonical prompt files in this repository. + +If `prompts/` must be created later, every file must be a product-local draft adapter and include this frontmatter: + +```yaml +--- +authority: product_local_adapter +status: draft +canonical_source: pending_ccpe +not_canonical_ccpe_prompt: true +--- +``` + +Adapter files may reference pending CCPE contracts but cannot claim canonical status. + +## Routing Rule + +When M2 needs canonical prompt/spec/rubric/protocol authority, write a CCPE request under: + +```text +requirements/ccpe/ +``` + +When M2 needs reusable deterministic automation, write a skills-vault request under: + +```text +requirements/skills-vault/ +``` + +Do not solve supplier-owned capabilities locally unless the owner explicitly changes the boundary. + +## Forbidden During M2a + +M2a must not: + +- create canonical prompts; +- create `ccra_review_bundle/`; +- start Local CCRA; +- import old prompt packs from the previous project; +- implement a full selector; +- implement reusable validators locally; +- expand the model set. + +## Revised M2a Plan + +Goal: correct M2 authority boundaries before any prompt/runtime skeleton work. + +Deliverables: + +- `docs/PROMPT_AUTHORITY_BOUNDARY.md`; +- M2 boundary decisions in `docs/DECISIONS.md`; +- CCPE request for cognitive runtime prompt contracts; +- skills-vault request for deterministic adapter/run manifest tooling if needed; +- request-channel READMEs. + +Acceptance: + +- no `prompts/` directory is required; +- no canonical CCPE artifact is copied into this repository; +- any missing canonical prompt/spec is represented as a supplier request; +- any reusable deterministic tool is represented as a supplier request. + +## Revised M2b Plan + +Goal: create only product-local runtime material after M2a boundaries are accepted. + +Allowed M2b work: + +1. Create product-local draft adapters only if the owner confirms adapter work should start before CCPE supplies canonical Prompt Cards. +2. Use the mandatory adapter frontmatter from this document. +3. Define product-local run input/output shapes for: + - intake/value assessment; + - QPI; + - lens orchestration; + - Intellectual Archaeology; + - synthesis/calibration; + - reader translation; + - feedback/asset decision. +4. Run one minimal manual sample only after adapters and run format are clearly product-local. +5. Record the run as product evidence, not as review evidence. + +M2b must still not start Local CCRA. Local CCRA waits until boundary correction and one minimal manual sample are complete. diff --git a/requirements/ccpe/2026-06-20-m2-cognitive-runtime-prompt-contracts.md b/requirements/ccpe/2026-06-20-m2-cognitive-runtime-prompt-contracts.md new file mode 100644 index 0000000..9de036a --- /dev/null +++ b/requirements/ccpe/2026-06-20-m2-cognitive-runtime-prompt-contracts.md @@ -0,0 +1,156 @@ +# CCPE Request: M2 Cognitive Runtime Prompt Contracts + +Date: 2026-06-20 + +Client repository: + +```text +C:\Users\wangq\Documents\Codex\work-projects\Cognitive-OS-Wantsong +``` + +Requested supplier repository: + +```text +C:\Users\wangq\Documents\Codex\ccpe-system +``` + +## 1. Product Context + +`Cognitive-OS-Wantsong` is a new product-runtime repository for a lightweight cognitive-processing flow. + +The product repo currently owns M0-M1 boundary docs, QPI and Intellectual Archaeology model cards, and the model registry. + +The next planned phase is M2, but the GPT-authored M2 plan proposed prompt files without separating: + +- product-local runtime adapters; +- canonical Prompt Cards; +- Agent Specs; +- Runtime Specs; +- model-call protocols; +- evaluation rubrics; +- Local CCRA review protocol. + +This request asks CCPE to own the canonical contract layer before the product repo creates any canonical-looking prompt assets. + +## 2. Needed Capability + +CCPE should define the canonical prompt/runtime contract set for the Cognitive-OS M2 runtime. + +Needed canonical artifacts: + +1. Prompt Card contract for Intake / Value Assessment. +2. Prompt Card contract for QPI routing. +3. Prompt Card contract for Lens Orchestration. +4. Prompt Card contract for Intellectual Archaeology runner. +5. Prompt Card contract for Synthesis and Calibration. +6. Prompt Card contract for Reader Translation. +7. Prompt Card contract for Feedback and Asset Decision. +8. Runtime Spec for chaining these cards in a manual or semi-automated run. +9. Model-call protocol describing how product-local adapters may invoke or reference canonical prompt contracts. +10. Evaluation rubric for judging internal trace and reader-facing output quality. +11. Boundary rule for when Local CCRA can review the milestone. + +## 3. Why This Belongs In CCPE + +This work defines canonical prompt authority, runtime protocol, evaluation rubric, and review criteria. + +Those are CCPE responsibilities. The product repo should only hold product-local adapters, sample runs, local output formats, and supplier requests. + +## 4. Current Blocked Task + +The product repo should not create canonical prompt files for M2 until the CCPE contract exists. + +Allowed interim product work: + +- document prompt authority boundaries; +- create product-local draft adapters only with explicit adapter frontmatter; +- run a minimal manual sample only after adapter status is unambiguous. + +Blocked canonical work: + +```text +prompts/intake-value-assessor.md +prompts/qpi-router.md +prompts/lens-orchestrator.md +prompts/intellectual-archaeology-runner.md +prompts/synthesis-engine.md +prompts/reader-translation-engine.md +prompts/calibration-checker.md +``` + +These filenames may exist later only as product-local adapters unless CCPE explicitly supplies canonical placement rules. + +## 5. Input Files And Source Paths + +Product repo files: + +```text +README.md +AGENTS.md +PROJECTS.md +docs/PROJECT_BRIEF.md +docs/COGNITIVE_WORKFLOW_V0.md +docs/MODEL_MANAGEMENT_V0.md +docs/MODEL_ORCHESTRATION_V0.md +docs/READER_TRANSLATION_LAYER_V0.md +docs/LOCAL_CCRA_MILESTONE_PROTOCOL.md +docs/ASSET_MIGRATION_POLICY.md +docs/DECISIONS.md +models/qpi.md +models/intellectual-archaeology.md +models/model-registry.json +``` + +Planning/source references: + +```text +C:\Users\wangq\Documents\Codex\knowledge-vault\work\internal\强哥的思想宇宙\Cognitive-OS-Wantsong项目总计划 v0.1.md +C:\Users\wangq\Documents\Codex\work-projects\the-mindscape-of-bro-tsong\PROJECTS.md +C:\Users\wangq\Documents\Codex\work-projects\the-mindscape-of-bro-tsong\requirements\ +``` + +## 6. Expected Output Artifacts + +CCPE should provide or register: + +- canonical Prompt Card templates or concrete cards for the seven M2 runtime steps; +- a Runtime Spec for manual chaining and later automation; +- model-call protocol covering model registry lookup, primary/support role limits, and reader translation handoff; +- evaluation rubric for: + - QPI routing quality; + - Intellectual Archaeology depth quality; + - synthesis/calibration quality; + - reader translation quality; + - evidence/action-boundary quality; +- Local CCRA milestone-readiness criteria for after the first minimal manual sample. + +## 7. Required Human Gate + +Owner approval is required before: + +- treating any product-local adapter as canonical; +- copying any CCPE-supplied artifact into this product repo; +- starting Local CCRA; +- expanding beyond QPI and Intellectual Archaeology; +- turning M2 output into stable model governance. + +## 8. Prohibited Behavior + +CCPE output should not require this product repo to: + +- store canonical Prompt Cards as local truth; +- create a full selector; +- create a full regression suite; +- start Local CCRA before a minimal manual sample; +- recreate old review-bundle or Round Conductor workflows; +- use `lite` suffixes for model filenames or model IDs. + +## 9. Urgency And Dependency + +Urgency: high before M2 prompt work. + +Dependency: + +- M2a boundary correction can complete immediately in this product repo. +- M2b product-local adapter work can proceed only if adapter authority is clearly marked. +- Canonical prompt/runtime authority waits for CCPE. diff --git a/requirements/ccpe/README.md b/requirements/ccpe/README.md new file mode 100644 index 0000000..9a641d1 --- /dev/null +++ b/requirements/ccpe/README.md @@ -0,0 +1,45 @@ +# CCPE Capability Requests + +This directory holds supplier requests for capabilities that should be designed or maintained in: + +```text +C:\Users\wangq\Documents\Codex\ccpe-system +``` + +Use this channel when `Cognitive-OS-Wantsong` needs: + +- canonical Prompt Card; +- Agent Spec; +- Runtime Spec; +- model-call protocol; +- evaluation rubric; +- Local CCRA canonical review protocol; +- lifecycle or milestone authority rule; +- integration registration guidance. + +Do not place canonical CCPE artifacts in this repository. Store the client request here, then let `ccpe-system` own the supplied artifact. + +## Request Format + +Use one Markdown file per request. + +Recommended filename: + +```text +YYYY-MM-DD-short-request-name.md +``` + +Each request should include: + +- product context; +- needed capability; +- why this belongs in CCPE; +- current blocked task; +- input files or source paths; +- expected output artifact; +- required human gate; +- urgency and dependency. + +## Pause Rule + +If product work depends on a missing CCPE capability, write the request here and pause the dependent step unless the owner explicitly says to solve it inside this repository. diff --git a/requirements/skills-vault/2026-06-20-m2-runtime-run-logger-and-adapter-manifest-checker.md b/requirements/skills-vault/2026-06-20-m2-runtime-run-logger-and-adapter-manifest-checker.md new file mode 100644 index 0000000..7c68d2f --- /dev/null +++ b/requirements/skills-vault/2026-06-20-m2-runtime-run-logger-and-adapter-manifest-checker.md @@ -0,0 +1,121 @@ +# Skills-Vault Request: M2 Runtime Run Logger And Adapter Manifest Checker + +Date: 2026-06-20 + +Client repository: + +```text +C:\Users\wangq\Documents\Codex\work-projects\Cognitive-OS-Wantsong +``` + +Requested supplier repository: + +```text +C:\Users\wangq\Documents\Codex\skills-vault +``` + +## 1. Product Context + +`Cognitive-OS-Wantsong` will enter M2 after boundary correction. + +M2 should create product-local runtime adapters and eventually run one minimal manual sample. It must not create canonical CCPE prompt artifacts. + +This creates a deterministic validation need: adapter files and run records should be checked for authority metadata, manifest completeness, and path consistency without judging model quality. + +## 2. Repeated Action Or Friction Point + +Likely repeated checks: + +- product-local adapter frontmatter includes the required authority fields; +- no adapter claims canonical CCPE prompt status; +- run manifest lists all runtime steps and output files; +- run log paths exist; +- sample input, internal trace, reader output, and owner note are cross-linked consistently; +- Local CCRA is not started before the required milestone conditions. + +These checks are deterministic. They should not become product repo scripts unless the owner explicitly asks. + +## 3. Proposed Skill Name + +Candidate name: + +```text +cognitive-runtime-manifest-checker +``` + +Alternative names: + +```text +prompt-adapter-boundary-checker +runtime-run-log-validator +``` + +## 4. Inputs + +The Skill should accept: + +- repository root; +- optional adapter directory path; +- optional run directory path; +- expected authority policy. + +Expected adapter frontmatter policy: + +```yaml +authority: product_local_adapter +status: draft +canonical_source: pending_ccpe +not_canonical_ccpe_prompt: true +``` + +## 5. Expected Outputs + +The Skill should produce: + +- a PASS / FAIL summary; +- a machine-readable JSON report; +- finding list with file paths and line numbers when possible. + +Suggested report fields: + +```json +{ + "status": "PASS", + "adapter_files_scanned": 0, + "run_files_scanned": 0, + "blocking_findings": [], + "warnings": [] +} +``` + +## 6. Safety Boundaries + +- Read-only by default. +- No content generation. +- No prompt rewriting. +- No model-quality judgment. +- No Local CCRA judgment. +- No file edits unless a future explicit `--write-report` option is requested. + +## 7. Validation Expectation + +Minimum validation cases: + +- adapter with correct frontmatter passes; +- adapter missing `not_canonical_ccpe_prompt: true` fails; +- adapter claiming `authority: ccpe_canonical` fails inside the product repo; +- missing run output path fails; +- Local CCRA directory present before the allowed milestone warns or fails according to config; +- JSON report is stable and machine-readable. + +## 8. Why This Should Become Reusable Automation + +The operation is deterministic metadata and manifest checking. It does not depend on QPI content quality or Intellectual Archaeology reasoning. + +It is likely reusable for other product repos that consume CCPE canonical prompt contracts while keeping product-local adapters separate. + +## 9. Current Blocked Task + +This is not blocking M2a boundary correction. + +It may become useful before M2b manual sample execution, especially if multiple adapter and run files are created. diff --git a/requirements/skills-vault/README.md b/requirements/skills-vault/README.md new file mode 100644 index 0000000..958dad2 --- /dev/null +++ b/requirements/skills-vault/README.md @@ -0,0 +1,43 @@ +# Skills-Vault Capability Requests + +This directory holds supplier requests for deterministic automation that should be implemented or maintained in: + +```text +C:\Users\wangq\Documents\Codex\skills-vault +``` + +Use this channel when `Cognitive-OS-Wantsong` needs: + +- prompt adapter validator; +- run logger; +- manifest checker; +- review context builder; +- deterministic file transformer; +- reusable CLI or installable Skill. + +Do not place reusable automation Skill source in this repository. Store the client request here, then let `skills-vault` own implementation source if the need becomes stable and reusable. + +## Request Format + +Use one Markdown file per request. + +Recommended filename: + +```text +YYYY-MM-DD-short-request-name.md +``` + +Each request should include: + +- product context; +- repeated action or friction point; +- input files and expected outputs; +- safety boundaries; +- whether file edits are allowed; +- validation expectation; +- why this should become reusable automation; +- current blocked task. + +## Pause Rule + +If product work depends on a missing reusable automation capability, write the request here and pause the dependent step unless the owner explicitly says to solve it inside this repository.