feat: Phase 2 Wave 1 — protocol slimming, answer-first, progressive retrieval, pinned core memory by fasdwcx · Pull Request #5 · relytcloud/engram

docs(spec): Phase 2 optimization loop design (Wave 1: R1-R5)

fasdwcx

committed

docs(plan): Phase 2 Wave 1 implementation plan (R1-R5 optimization loop)

fasdwcx

committed

feat(eval): wave1 guardrails — distractor ratio, compare suite, task filter, mem_* audit log

fasdwcx

committed

fix(eval): guard partial L3 scorecards, tolerate trailing task csv, test audit passthrough, document audit env

fasdwcx

committed

feat(protocol): revive slim protocol-mode for fork versions and add --set

fasdwcx

committed

fix(protocol): correct slim floor version in setup help text

fasdwcx

committed

feat(protocol): compact slim protocol text (~850 tok) behind protocol-mode

fasdwcx

committed

fix(protocol): setup help text reflects codex slim support

fasdwcx

committed

feat(protocol): trim MCP server instructions to ~325 tok; SKILL absorbs conflict-loop detail

fasdwcx

committed

fix(protocol): qualify deferred-tool list preloading and pin the save rule in the budget test

fasdwcx

committed

feat(eval): wave1 round 1 (R1 protocol slimming) scorecards

fasdwcx

committed

feat(protocol): answer-first save rules everywhere + slim survives compaction

fasdwcx

committed

fix(protocol): answer-first rules in all remaining protocol copies; pin full-branch rules in golden test

fasdwcx

committed

fix(protocol): answer-first mem_save tool description, guard post-compaction full branch, align docs

fasdwcx

committed

fix(protocol): front-matter-scoped answer-first guard for SKILL descriptions

fasdwcx

committed

feat(eval): wave1 round 2 (R2 answer-first) scorecards

fasdwcx

committed

feat(retrieval): mem_search returns budgeted reference-first index; L1 measures agent-visible payload

fasdwcx

committed

fix(retrieval): budget gates structured entries, L1 measures full payload, summary sentence guard

fasdwcx

committed

docs(retrieval): document SearchPayloadTokens as a floor measurement

fasdwcx

committed

feat(retrieval): layered mem_context with 400-token budget (both backends)

fasdwcx

committed

fix(retrieval): pin context drop-order/cap contracts, document tail and prompts posture

fasdwcx

committed

feat(eval): wave1 round 3 (R3 progressive retrieval) scorecards

fasdwcx

committed

feat(memory): pinned core-memory block with 1KB cap and stable ordering

fasdwcx

committed

fix(memory): never emit a zero pin-cap marker; truncate titles in context lines

fasdwcx

committed

feat(eval): wave1 round 4 (R4 pinned core memory) scorecards

fasdwcx

committed

docs(proposals): MemoryLake server asks from wave1 evidence

fasdwcx

committed

docs(proposals): scope validity claim to superseded share and state the wire envelope convention

fasdwcx

committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

feat: Phase 2 Wave 1 — protocol slimming, answer-first, progressive retrieval, pinned core memory - #5

feat: Phase 2 Wave 1 — protocol slimming, answer-first, progressive retrieval, pinned core memory#5
fasdwcx merged 29 commits into
mainfrom
feat/phase2-wave1

Commits on Jul 27, 2026

docs(spec): Phase 2 optimization loop design (Wave 1: R1-R5)

docs(plan): Phase 2 Wave 1 implementation plan (R1-R5 optimization loop)

feat(eval): wave1 guardrails — distractor ratio, compare suite, task filter, mem_* audit log

fix(eval): guard partial L3 scorecards, tolerate trailing task csv, test audit passthrough, document audit env

feat(protocol): revive slim protocol-mode for fork versions and add --set

fix(protocol): correct slim floor version in setup help text

feat(protocol): compact slim protocol text (~850 tok) behind protocol-mode

fix(protocol): setup help text reflects codex slim support

feat(protocol): trim MCP server instructions to ~325 tok; SKILL absorbs conflict-loop detail

fix(protocol): qualify deferred-tool list preloading and pin the save rule in the budget test

feat(eval): wave1 round 1 (R1 protocol slimming) scorecards

feat(protocol): answer-first save rules everywhere + slim survives compaction

fix(protocol): answer-first rules in all remaining protocol copies; pin full-branch rules in golden test

fix(protocol): answer-first mem_save tool description, guard post-compaction full branch, align docs

fix(protocol): front-matter-scoped answer-first guard for SKILL descriptions

feat(eval): wave1 round 2 (R2 answer-first) scorecards

feat(retrieval): mem_search returns budgeted reference-first index; L1 measures agent-visible payload

fix(retrieval): budget gates structured entries, L1 measures full payload, summary sentence guard

docs(retrieval): document SearchPayloadTokens as a floor measurement

feat(retrieval): layered mem_context with 400-token budget (both backends)

fix(retrieval): pin context drop-order/cap contracts, document tail and prompts posture

feat(eval): wave1 round 3 (R3 progressive retrieval) scorecards

feat(memory): pinned core-memory block with 1KB cap and stable ordering

fix(memory): never emit a zero pin-cap marker; truncate titles in context lines

feat(eval): wave1 round 4 (R4 pinned core memory) scorecards

docs(proposals): MemoryLake server asks from wave1 evidence

docs(proposals): scope validity claim to superseded share and state the wire envelope convention

Commits on Jul 28, 2026

feat(eval): wave1 acceptance run and report

fix(eval): correct auth-artifact-contaminated L3 anchors on both sides; honest wave1 report