Plan · NovelForge 7.1 Adaptive Runtime

Strategy

Implement 7.1 as a deterministic shell around optional host/model capabilities. The shell owns identity, state transitions, fingerprints, provenance, rights/storage policy, idempotency, queue construction, promotion prerequisites, bundle materialization and verification. Hosts/models perform only the work they actually have capability for.

Phase 1 · Capability Contract

Create harness/runtime_capabilities.py with:

• novelforge_host_capabilities_v1 schema;
• local probe for executable/runtime facts that can be proven locally;
• normalization for externally supplied Chat/MCP/Web/GitHub capability declarations;
• requirement resolution that never selects undeclared/unavailable capability;
• usage/cost/user-interaction metadata;
• self-test.

Add paired runtime-capability documentation and wire it into Runtime Routing.

Phase 2 · Discovery Runtime

Create corpus/discovery_runtime.py with:

• typed dispatch plan built from Corpus Scout requests + host capabilities;
• typed discovery result validation;
• source/tool/channel provenance binding;
• evidence fingerprint;
• deterministic rights/storage validation using rights_gate.py;
• dedupe by source locator/content fingerprint;
• diversity summary by work/source/channel;
• self-test.

Upgrade corpus_scout.py request schema so each requested channel declares a capability requirement rather than assuming host access.

Phase 3 · Learning Cycle

Create learning/learning_cycle.py with a SQLite-backed state machine in the same learning DB, but separate cycle tables. It will:

• start from an existing Corpus gap/hypothesis;
• persist cycle identity/state/version;
• register discovery queues/results;
• register analysis/eval queues/results by fingerprint;
• enforce legal transitions;
• enforce logical consume-once;
• resume from durable state;
• never grant Canon/Framework-write authority;
• self-test crash/retry/idempotency semantics.

Phase 4 · Semantic Learning Work

Create learning/learning_eval.py to package:

• corpus_analyze semantic jobs from verified discovery evidence;
• preference_distill or external_review jobs when scope requires;
• eval_judge jobs for capability/regression evidence;
• typed output contracts for mechanism, counterexample, boundary, evidence refs and confidence.

Use the existing semantic fingerprint function; execution lineage remains outside semantic fingerprint.

No hidden gold or expected labels enter worker packets.

Phase 5 · Promotion Gate

Create learning/promotion_gate.py.

Rules:

• one_off: never durable promotion;
• project: may become project proposal/active preference only with project authority, not from Generic Framework code;
• user_taste: requires explicit/repeated human evidence + contradiction review + eval evidence; gate can mark ready but not persist private taste to Generic source;
• general_craft: requires cross-work evidence, counterexample/profile boundary, capability+regression eval, target version, rollback ref and green CI evidence.

The gate outputs evidence completeness and blockers. It never edits Framework behavior.

Phase 6 · Immutable Bundle

Create release/build_framework_bundle.py and paired docs.

Bundle rules:

• sorted paths;
• deterministic tar metadata (mtime=0, fixed uid/gid/mode normalization);
• exclude .git, generated artifacts, runtime DBs, specs and bundle metadata from runtime bundle by policy;
• include Generic Core/Surface/Harness/Learning/Corpus/Evals/SDK/docs/runtime bootstrap;
• content manifest with per-file SHA-256;
• overall bundle SHA-256;
• verify command;
• self-test reproduces identical fingerprint and detects tamper.

Create optional novelforge-release-bundle.yml to build/verify/upload the immutable artifact. Normal CI validates the builder but does not publish a release or spend model usage.

Phase 7 · CLI / CI / Maintenance

Upgrade novelforge.py:

• framework version 7.1.0;
• capabilities router;
• learning-cycle router;
• learning-gate router;
• corpus discovery router;
• bundle router;
• self-test covers all deterministic modules.

Upgrade reusable CI:

• compile all modules;
• run new self-tests;
• assert no model execution;
• build bundle twice and compare fingerprints;
• inspect adaptive queue invariants.

Upgrade weekly maintenance:

• probe only available deterministic/local capabilities;
• create learning/discovery work queues;
• report pending capability requirements;
• never claim unexecuted Web/model work;
• never auto-promote.

Phase 8 · Docs / Manifest / Version

Update:

• HARNESS_MANIFEST.yaml → 7.1.0;
• SKILL.md, SKILL.en.md, SKILL.zh-CN.md;
• Harness/Runtime Routing/Self-Improvement docs;
• adaptive-learning docs;
• Corpus docs/policy;
• integrations docs;
• Project SDK docs;
• README + CHANGELOG;
• project_sdk.py default/minimum version behavior.

Record external mechanism evidence from official OpenAI Agents SDK, LangGraph, MCP and Google ADK docs without turning them into dependencies.

Phase 9 · Framework Release Verification

1. Push implementation on main.
2. Wait for final NovelForge CI.
3. If failed, repair the owning mechanism; do not bless a failing commit.
4. Once green, run deterministic bundle workflow/build and capture fingerprint.
5. Commit bundle attestation metadata if needed; bundle metadata itself remains outside fingerprint input so no circular hash.
6. Re-run CI and verify final framework HEAD.

Phase 10 · Chinatown Consumer Upgrade

After final 7.1 framework HEAD is green:

• update novelforge.toml minimum framework version to 7.1.0;
• update novelforge.lock.json exact commit + bundle fingerprint;
• update framework attestation;
• update Project bootstrap docs only where 7.1 capability/bundle semantics matter;
• strengthen project validator to require bundle fingerprint for 7.1 locks;
• run Chinatown Project CI;
• verify no Canon/active-story mutation.

Rollback

Framework rollback: de05666cc4eae13f09868d87659e76f2aa524314.

Consumer rollback: restore the last 7.0 lock/attestation and project bootstrap commit if the 7.1 consumer gate fails. No Canon migration is part of this release.