From 311c95c17c545aea327afffccb5476db46e950ad Mon Sep 17 00:00:00 2001 From: RightL Date: Wed, 30 Sep 2026 14:27:47 +0800 Subject: [PATCH 1/5] prompt: clarify missing context submissions --- rightmemory/mcp.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/rightmemory/mcp.py b/rightmemory/mcp.py index 3532d61..f15c0e8 100644 --- a/rightmemory/mcp.py +++ b/rightmemory/mcp.py @@ -56,6 +56,8 @@ SUBMIT_UPDATE_DESCRIPTION = """\ Submit durable Memory or Agent Correction evidence when omitting it would likely cause poorer future decisions, substantial rediscovery, or loss of useful context. +Include concrete findings of missing or ambiguous context that meet this bar, even +when the answer is not yet known. Pursuit is read-only to Update. Explicit map edits belong in the human editor or the maintain-pursuit-map workflow, not in an Update submission. From 9abea119bf81a1941a7c58c82903d1d332894866 Mon Sep 17 00:00:00 2001 From: RightL Date: Wed, 30 Sep 2026 14:28:01 +0800 Subject: [PATCH 2/5] prompt: preserve unresolved context gaps --- rightmemory/prompts/update.md | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/rightmemory/prompts/update.md b/rightmemory/prompts/update.md index 078ddc4..7add04b 100644 --- a/rightmemory/prompts/update.md +++ b/rightmemory/prompts/update.md @@ -13,6 +13,7 @@ Reconcile the supplied candidates into durable Memory and Agent Corrections. Can - If a candidate mixes durable evidence with a possible Pursuit map change, judge the durable portion normally and report the map portion as skipped. Do not infer or apply Pursuit changes. - Do not preserve candidate ids, submission labels, or operational event names unless they independently matter to the user. - Revalidate claims that may have become stale instead of copying candidates as fact. +- Preserve consequential context gaps or ambiguity found in the candidates or related state as Open Context Questions when available evidence cannot resolve them. ## Plan The Edit @@ -30,7 +31,7 @@ Reconcile the supplied candidates into durable Memory and Agent Corrections. Can - Preserve unrelated valid content while using enough scope to keep the complete affected graph coherent. - Validate the complete graph before finishing. - If state changed, stage and commit only allowed RightMemory files touched by the update. -- If no candidate survives reconciliation and no encountered violation requires repair, make no commit. +- If reconciliation and permitted repairs leave state unchanged, make no commit. ## Final Reply From 70751b9f151c9a0040f6e84b1ad4d48ed30723a0 Mon Sep 17 00:00:00 2001 From: RightL Date: Wed, 30 Sep 2026 14:28:03 +0800 Subject: [PATCH 3/5] skill: preserve open context questions --- skills/maintain-rightmemory/SKILL.md | 1 + 1 file changed, 1 insertion(+) diff --git a/skills/maintain-rightmemory/SKILL.md b/skills/maintain-rightmemory/SKILL.md index 0368395..5d07c83 100644 --- a/skills/maintain-rightmemory/SKILL.md +++ b/skills/maintain-rightmemory/SKILL.md @@ -7,6 +7,7 @@ description: "Use when the user explicitly asks to inspect or maintain RightMemo - Run `rightmemory validate --root ` before reading RightMemory state. Never follow or read a path it reports as non-regular; treat other findings as maintenance evidence. - Read the package-owned schema and Memory rules with `rightmemory reference schema` and `rightmemory reference memory`, then read available regular `/MEMORY.md`, relevant reachable backings, and relevant Git history or `update_records/*.json`. Read Pursuit as context for Memory meaning and references, not as content to reorganize. Before judging corresponding content, read `rightmemory reference pursuit`, `rightmemory reference agent-correction`, `rightmemory reference edit-correction`, or `rightmemory reference shared-view`; never edit immutable update records. Inspect the affected reachable state for concrete organization, consolidation, and cleanup that would improve future use. - During cleanup, delete trivial items instead of compacting them. +- Preserve consequential context gaps or ambiguity found within the requested scope as Open Context Questions; in read-only reviews, report them without editing. - An explicit request to change these surfaces authorizes the requested maintenance. Read-only inspection does not authorize edits. Resolve ordinary local details from the current state, and ask only when a missing target or materially different interpretation prevents a safe change; do not add a routine preview-and-approval ceremony. - Before writing, require the active root to be clean, capture its branch and HEAD, and create a dedicated temporary Git worktree and branch from that HEAD. Do not stash, discard, or absorb unrelated changes. - Apply the requested changes in the worktree to Memory, its F#/M#/S# backings, Agent Corrections, RightMemory Edit Feedback, and supporting references. Keep Agent Corrections separate from root `corrections.md`. Preserve ids referenced from Pursuit and leave the Pursuit document tree unchanged. Explicit map changes belong to the separate `maintain-pursuit-map` workflow; do not invoke it or queue map work from this skill. From a128b08e473e33172a123d01be90cc356683b82d Mon Sep 17 00:00:00 2001 From: RightL Date: Wed, 30 Sep 2026 14:28:07 +0800 Subject: [PATCH 4/5] skill: propose consequential context gaps --- skills/rightmemory-orchestrator-cli/SKILL.md | 1 + 1 file changed, 1 insertion(+) diff --git a/skills/rightmemory-orchestrator-cli/SKILL.md b/skills/rightmemory-orchestrator-cli/SKILL.md index 232aa73..6480f8e 100644 --- a/skills/rightmemory-orchestrator-cli/SKILL.md +++ b/skills/rightmemory-orchestrator-cli/SKILL.md @@ -31,6 +31,7 @@ Choose one stable session id for the conversation and reuse it for every RightMe - Propose only when omitting the evidence would likely: - cause poorer future decisions or substantial rediscovery (**Memory**); or - make repetition of a settled, reusable failure pattern more likely (**Agent Corrections**). +- Propose concrete findings of missing or ambiguous context that meet this bar, even when the answer is not yet known. - Do not propose transient progress, routine task results, unfinished work by itself, or implementation detail already adequately preserved in project-local artifacts. - Name the apparent module and briefly explain why the evidence passes this bar. Do not propose final stored wording, ids, or edits. - Present the proposal once the evidence is clear and the conversation reaches a natural boundary. Completion is not required. If nothing passes the bar, say nothing about updating RightMemory. From 27975df613417fe2587f6489d3976effcdd2be1d Mon Sep 17 00:00:00 2001 From: RightL Date: Wed, 30 Sep 2026 14:28:09 +0800 Subject: [PATCH 5/5] skill: submit consequential context gaps --- skills/rightmemory-auto-orchestrator-cli/SKILL.md | 1 + 1 file changed, 1 insertion(+) diff --git a/skills/rightmemory-auto-orchestrator-cli/SKILL.md b/skills/rightmemory-auto-orchestrator-cli/SKILL.md index 5be0182..361839d 100644 --- a/skills/rightmemory-auto-orchestrator-cli/SKILL.md +++ b/skills/rightmemory-auto-orchestrator-cli/SKILL.md @@ -26,6 +26,7 @@ Choose one stable session id for the conversation and reuse it for every RightMe ## Submit Updates - Submit automatic Memory evidence when omitting it would likely cause poorer future decisions or substantial rediscovery. +- Include concrete findings of missing or ambiguous context that meet this bar, even when the answer is not yet known. - Do not submit transient progress, routine task results, unfinished work by itself, or implementation detail already adequately preserved in project-local artifacts. - Submit once the durable evidence is clear and the work reaches a natural boundary. Completion is not required. If nothing passes the bar, submit nothing; task progress and new ideas do not trigger a map-change suggestion. - Combine related evidence due at the same boundary into one candidate. State what happened, what is true now, and why it may matter; do not prescribe final stored wording, ids, classification, placement, or edits.