Release v1.6.0 stage continuity recall guards

Author: Template Maintainer

.agents/skills/agent-orchestration/SKILL.md

@@ -19,6 +19,10 @@ The user may grant project-level standing permission for the agent to decide whe
 
 Before drafting any formal document, use `dissertation-source-first-gate`. Extract known factual fields from source files. Do not invent names, emails, dates, institutional contacts, supervisor/PI/client details, signatures, participant facts, datasets, results, administrative requirements, journal/funder/client requirements, marking criteria, grade-band standards, LMS requirements, deadlines, word counts, or submission rules. If a field is not found in the source material, mark it `TO CONFIRM`.
 
+Before substantive cross-stage work, apply `research-wiki/STAGE_CONTINUITY_PROTOCOL.md` when both conditions hold: (A) the task changes, designs, restates, translates, formalises, or produces a high-risk deliverable class such as method plan, compliance material, fieldwork instrument, analysis plan, stakeholder-facing decision memo, or formal section draft; and (B) `research-wiki/STAGE_GRAPH.md` lists upstream dependencies or the task references prior decisions, accepted outputs, source-of-record files, source maps, or checkpoints. Citation-only edits, formatting/layout fixes, typo repair, file moves, bookkeeping, and source summaries without design/formal-claim output do not trigger the gate.
+
+Use `scripts/stage_recall_policy.py` or the runtime preflight `recall_decision` as the token-aware recall controller. Tier 0 means no project recall; Tier 1 anchor scan; Tier 2 pointer lookup; Tier 3 targeted Stage Continuity Capsule; Tier 4 full upstream audit or pause. This controller manages context budget only. It never overrides Stage Continuity A+B, source-first, compliance, citation, privacy, document-quality, or delivery gates. If the user says to skip upstream checks for a triggered deliverable, surface the omitted dependency first and record an override risk only after explicit user acceptance.
+
 Before answering rubric, grade-band, marking-criteria, journal/funder/client requirement, LMS requirement, word-count, deadline, or submission-rule questions, use the strongest available project requirement source. Use `university-guidance/RUBRIC_EVIDENCE_GATE.md` for assessed academic work. Distinguish official original text, local summary, inference, and evidence-insufficient status.
 
 Before delivering formal documents or important project notes, use `dissertation-document-quality-gate` at the appropriate level.
@@ -37,10 +41,13 @@ Use this quick sequence at the start of every Production Window turn:
 
 1. Classify the user's current task by mode and task type.
 2. Select the smallest useful skill set from the classification table.
-3. Open the relevant `SKILL.md` files before acting on any non-trivial task.
-4. Add source-first, cognitive-framework, self-review, argument-spine, style-memory, document-quality, compliance, project-delivery, and context-continuity gates when the task involves formal or stakeholder-facing output.
-5. Re-route if the user changes the task during the turn.
-6. Record the routing in `research-wiki/PRODUCTION_RUN_REGISTER.md` for substantial Production tasks.
+3. Compute or consume the Token-Aware Recall tier and state the minimum recall scope for substantial work.
+4. Check whether Stage Continuity A+B is triggered. If yes, read `research-wiki/STAGE_GRAPH.md`, write a Stage Continuity Capsule, and run `scripts/stage_continuity_capsule_check.py` when available.
+5. Open the relevant `SKILL.md` files before acting on any non-trivial task.
+6. Add source-first, cognitive-framework, self-review, argument-spine, style-memory, document-quality, compliance, project-delivery, and context-continuity gates when the task involves formal or stakeholder-facing output.
+7. Re-route and recompute recall if the user changes the task during the turn.
+8. Before formal delivery, confirm the final artifact still matches the latest recall tier, source map, and gates.
+9. Record the routing in `research-wiki/PRODUCTION_RUN_REGISTER.md` for substantial Production tasks.
 
 Use this receipt wording:
 
@@ -53,6 +60,8 @@ Skill routing:
 - Subagent decision:
 - Gates required:
 - Gates completed:
+- Recall Decision Note:
+- Stage Continuity Gate triggered:
 ```
 
 If a substantial Production output has no skill-routing receipt, treat the task as incomplete for maintenance-audit purposes.
@@ -82,6 +91,7 @@ Classify the user's request into one or more categories:
 | LMS/module requirements | `dissertation-research-search-protocol`, `dissertation-research-wiki`, `dissertation-chapter-plan`, `dissertation-citation-audit` |
 | literature search / literature review | `dissertation-research-search-protocol`, `dissertation-learning-loop`, `dissertation-literature-review`, `cognitive-frameworks`, `dissertation-argument-spine`, `dissertation-citation-audit`, `academic-self-review-loop` when drafting formal synthesis |
 | research questions / methodology | `cognitive-frameworks`, `dissertation-argument-spine`, `dissertation-research-review`, `dissertation-chapter-plan`, `academic-self-review-loop` when drafting formal prose |
+| Stage Continuity triggered deliverable | `context-continuity`, `cognitive-frameworks`, `dissertation-source-first-gate`, `dissertation-argument-spine`, `dissertation-research-review` |
 | interview guide / data collection | `qualitative-theme-audit`, `responsible-ai-agent-audit`, `teacher-adoption-modeling` |
 | confirmed design-elicitation / co-design outputs | `codesign-output-synthesis`, `qualitative-theme-audit`, `ai-agent-design-spec` |
 | AI agent concept / prototype | `ai-agent-design-spec`, `active-learning-design-support`, `prototype-evaluation-audit` |

.agents/skills/context-continuity/SKILL.md

@@ -5,7 +5,7 @@ description: Maintain compact-ready checkpoints, task state summaries, source ma
 
 # Context Continuity
 
-Use this skill for long, multi-step, or easily interrupted dissertation tasks, especially when work may span many turns, involve multiple files, or risk losing source-grounded decisions.
+Use this skill for long, multi-step, or easily interrupted research tasks, especially when work may span many turns, involve multiple files, cross project stages, or risk losing source-grounded decisions.
 
 ## Purpose
 
@@ -21,8 +21,10 @@ This skill also carries the adapted ECC context-budget, save-session, and strate
    - output files
    - confirmed facts
    - `TO CONFIRM` fields
-2. During the task, update a compact checkpoint when major decisions or files change.
-3. Before final response, summarize:
+2. If `research-wiki/STAGE_CONTINUITY_PROTOCOL.md` triggers, read `research-wiki/STAGE_GRAPH.md` before substantive work and record a Stage Continuity Capsule.
+3. Use the runtime `recall_decision` or `scripts/stage_recall_policy.py` as the context-budget controller. Recompute when the task changes from discussion to formal output, from layout to content, from reading to design/method/analysis, or before formal delivery.
+4. During the task, update a compact checkpoint when major decisions or files change.
+5. Before final response, summarize:
    - what was done
    - what files changed
    - what evidence was used
@@ -59,6 +61,7 @@ Update `research-wiki/TASK_STATE.md` when:
 - a new source cluster or contextual source is added
 - a project rule or skill changes
 - a decision affects methodology, research questions, concept cards, data collection, or participant-facing materials
+- a Stage Continuity Capsule or Deep Reasoning Pass changes what later work must inherit
 - a long task is likely to be resumed later
 
 For substantial Production Window tasks, also update `research-wiki/PRODUCTION_RUN_REGISTER.md` with a run receipt. This lets the Maintenance Window compare the claimed skill routing, created files, gates performed, render artifacts, temporary files, and remaining risks.
@@ -106,6 +109,32 @@ Next action:
 
 For substantial Production tasks, also add a receipt to `research-wiki/PRODUCTION_RUN_REGISTER.md` using the current receipt fields. Do not rely on `TASK_STATE.md` alone for cross-window monitoring.
 
+## Stage Continuity Capsule
+
+Use `research-wiki/STAGE_CONTINUITY_PROTOCOL.md` when triggered. The capsule can live in a thinking checkpoint, design note, decision memo, or `research-wiki/TASK_STATE.md`.
+
+Minimum fields:
+
+```text
+Stage Continuity Capsule:
+- Current task/stage:
+- Trigger:
+- Stage graph nodes used:
+- Source-of-record files checked:
+- Inherited decisions:
+- Open confirmations / hard stops:
+- What may change:
+- What must not change without confirmation:
+- Next action boundary:
+```
+
+Content rules:
+
+- Use concrete local file paths.
+- Cite a stage node or source path for inherited decisions.
+- For high-risk deliverables, keep `What must not change without confirmation` non-empty.
+- Run `scripts/stage_continuity_capsule_check.py` when available.
+
 ## Where To Write
 
 Use:

AGENTS.md

@@ -85,6 +85,10 @@ Domain-specific skills are included as optional examples. Rename, edit, or remov
   - `python3 scripts/agent_runtime.py "<TASK>" --window Production --write --strict`
   - or `python3 scripts/agent_runtime.py "<TASK>" --window Maintenance --write --strict`
   If it returns `BLOCKED`, fix the missing file or gate before continuing.
+- For long-running projects, use `research-wiki/STAGE_GRAPH.md` and `research-wiki/STAGE_CONTINUITY_PROTOCOL.md` before a later-stage task changes, designs, restates, translates, formalises, or produces a high-risk deliverable with upstream dependencies. High-risk deliverables include proposal/brief material, method plans, compliance or ethics material, fieldwork instruments, concept cards/scenario stimuli, RQ-to-method mapping, analysis plans, stakeholder-facing decision memos, and formal chapter/section drafts.
+- Use the runtime `recall_decision` or run `python3 scripts/stage_recall_policy.py --task "<TASK>"` as the Token-Aware Recall controller. Tier 0=no project recall, Tier 1=anchor scan, Tier 2=pointer lookup, Tier 3=targeted Stage Continuity Capsule, Tier 4=full upstream audit or pause. This controller saves context but cannot override source-first, compliance, citation, privacy, document-quality, delivery, or Stage Continuity A+B gates.
+- If a user asks to skip an upstream check for a triggered deliverable, surface the omitted dependency first. Only after explicit user acceptance may the task proceed as an override risk; do not call the Stage Continuity Gate a pass.
+- For non-obvious route, method, instrument, analysis, or delivery decisions, write a concise Deep Reasoning Pass before drafting: decision under consideration, chosen direction and concrete trade-off accepted, rejected alternative only if genuinely considered, and what would change the decision. Do not expose private chain-of-thought.
 - Do not invent names, emails, supervisor/PI/client details, dates, funder/journal/client/institutional requirements, rubrics, citations, participant facts, datasets, results, or findings.
 - For formal drafting or editing, use `dissertation-source-first-gate`.
 - For substantial proposal, manuscript, report, grant, literature review, methodology, or stakeholder-facing writing, use source-first, then `material-passport`, then `academic-integrity-preflight`, then `cognitive-frameworks` before drafting.
@@ -143,34 +147,35 @@ Domain-specific skills are included as optional examples. Rename, edit, or remov
 2. Run `scripts/agent_runtime.py` for substantial tasks when local tools are available.
 3. Read `RESEARCH_PROJECT_BRIEF.md` if present; otherwise use `RESEARCH_PROJECT_BRIEF_TEMPLATE.md` and mark project facts `TO CONFIRM`.
 4. Read `PROJECT_AGENT_PREFERENCES.md` and relevant task-state files.
-5. Use source-first checks before formal writing.
-6. Use `material-passport` to package source, compliance/requirement, citation, and `TO CONFIRM` status before formal artifacts move forward.
-7. Use `academic-integrity-preflight` before major revision and again before delivery.
-8. Use `cognitive-frameworks` before major argument, gap, methodology, literature, proposal, manuscript, report, grant, or stakeholder-facing drafting.
-9. Use `academic-self-review-loop` before style polishing and document-quality review for formal prose.
-10. Use `authorial-voice-integrity` when the task involves AI-style prose, humanising, de-AI, detector framing, or AI-use disclosure.
-11. Use `style-fingerprint-gate` for formal prose before delivery when repeated contrast templates could become visible.
-12. Record required skill execution receipts for substantial formal tasks.
-13. Use the learning loop after useful reading or confirmed decisions.
-14. Use `knowledge-base/self-growing/` for controlled intake, growth queue triage, and compiled-wiki navigation.
-15. Use source-readiness checks before citation-heavy writing.
-16. Use compliance checks before ethics, privacy, funder, journal, client, or data-management claims.
-17. Use rubric or requirement evidence checks before grade-band, journal, funder, deadline, or word-count claims.
-18. Use `research-wiki/DOCUMENT_PIPELINE.md` for important Word/PDF/stakeholder-facing delivery.
-19. Use the project delivery review gate before formal document delivery.
-20. Use `formal-delivery-guard` before presenting formal artifacts as usable.
-21. Use relevant academic/professional style gates before delivering prose.
-22. Use document-quality gate before delivering formal outputs.
-23. Update `research-wiki/TASK_STATE.md` after substantial work.
-24. Record substantial Production work in `research-wiki/PRODUCTION_RUN_REGISTER.md` if that register is enabled.
-25. Use `brainstorming` for unclear, high-impact route decisions before drafting or system changes.
-26. Use `project-skill-creator-governance` and global `skill-creator` before adding or changing skills.
-27. Use `playwright-dissertation-browser` and global `playwright` for controlled browser automation.
-28. Use `markitdown` only after checking tool availability and privacy boundaries.
-29. Use `research-*` figure/writing skills only as optional quality layers after source, privacy, compliance, citation, and document gates.
-30. Use `scripts/claude_independent_review.py` for optional context-naive independent review when the artifact is safe to send to Claude Code.
-31. Use staged literature gap-watch automation only for candidate discovery unless the user confirms ingestion.
-32. Use `release-surface-verification` before saying a public GitHub release or template update is visible and ready for readers.
+5. Compute or consume the Token-Aware Recall tier; apply Stage Continuity A+B when triggered.
+6. Use source-first checks before formal writing.
+7. Use `material-passport` to package source, compliance/requirement, citation, and `TO CONFIRM` status before formal artifacts move forward.
+8. Use `academic-integrity-preflight` before major revision and again before delivery.
+9. Use `cognitive-frameworks` before major argument, gap, methodology, literature, proposal, manuscript, report, grant, or stakeholder-facing drafting.
+10. Use `academic-self-review-loop` before style polishing and document-quality review for formal prose.
+11. Use `authorial-voice-integrity` when the task involves AI-style prose, humanising, de-AI, detector framing, or AI-use disclosure.
+12. Use `style-fingerprint-gate` for formal prose before delivery when repeated contrast templates could become visible.
+13. Record required skill execution receipts for substantial formal tasks.
+14. Use the learning loop after useful reading or confirmed decisions.
+15. Use `knowledge-base/self-growing/` for controlled intake, growth queue triage, and compiled-wiki navigation.
+16. Use source-readiness checks before citation-heavy writing.
+17. Use compliance checks before ethics, privacy, funder, journal, client, or data-management claims.
+18. Use rubric or requirement evidence checks before grade-band, journal, funder, deadline, or word-count claims.
+19. Use `research-wiki/DOCUMENT_PIPELINE.md` for important Word/PDF/stakeholder-facing delivery.
+20. Use the project delivery review gate before formal document delivery.
+21. Use `formal-delivery-guard` before presenting formal artifacts as usable.
+22. Use relevant academic/professional style gates before delivering prose.
+23. Use document-quality gate before delivering formal outputs.
+24. Update `research-wiki/TASK_STATE.md` after substantial work.
+25. Record substantial Production work in `research-wiki/PRODUCTION_RUN_REGISTER.md` if that register is enabled.
+26. Use `brainstorming` for unclear, high-impact route decisions before drafting or system changes.
+27. Use `project-skill-creator-governance` and global `skill-creator` before adding or changing skills.
+28. Use `playwright-dissertation-browser` and global `playwright` for controlled browser automation.
+29. Use `markitdown` only after checking tool availability and privacy boundaries.
+30. Use `research-*` figure/writing skills only as optional quality layers after source, privacy, compliance, citation, and document gates.
+31. Use `scripts/claude_independent_review.py` for optional context-naive independent review when the artifact is safe to send to Claude Code.
+32. Use staged literature gap-watch automation only for candidate discovery unless the user confirms ingestion.
+33. Use `release-surface-verification` before saying a public GitHub release or template update is visible and ready for readers.
 
 ## Public Template Boundary
 

CHANGELOG.md

@@ -1,5 +1,34 @@
 # Changelog
 
+## v1.6.0 - Stage Continuity And Token-Aware Recall - 2026-06-10
+
+Status: long-running project continuity and context-budget reliability update.
+
+### Added
+
+- `research-wiki/STAGE_GRAPH.md` as a generic, user-customisable pointer map for upstream source-of-record dependencies.
+- `research-wiki/STAGE_CONTINUITY_PROTOCOL.md` to define Stage Continuity A+B triggers, non-triggers, user-skip override handling, Deep Reasoning Pass, and capsule requirements.
+- `scripts/stage_recall_policy.py` to compute deterministic recall tiers from task intent, target files, and change type.
+- `scripts/stage_continuity_capsule_check.py` to check capsule fields, concrete source paths, and confirmation boundaries.
+- Stage continuity unit tests and eval cases `STAGE-001`, `STAGE-002`, and `STAGE-003`.
+- Claude Code advisory review packet and review report for this public sync.
+
+### Changed
+
+- `scripts/agent_runtime.py` now emits a `recall_decision` for every preflight and adds Stage Continuity gates only when recall reaches Tier 3 or higher.
+- `agent-orchestration` now requires opening recall, mid-task recall recomputation, and pre-delivery recall reconciliation for substantial stage-sensitive work.
+- `context-continuity` now owns Stage Continuity Capsules and records what later work must inherit.
+- `AGENTS.md`, `PROJECT_AGENT_PREFERENCES.md`, `README.md`, and `README_CN.md` now document the workflow: opening recall prevents blind drafting, mid-task recall prevents drift, and delivery gates prevent packaging drift as a formal artifact.
+- Skill eval registry now reports 38 public checks, including three new Stage Continuity cases.
+
+### Boundary
+
+- Stage Graph rows are starter examples, not required private-project structure. A one-row graph is valid after customisation.
+- Token-Aware Recall is a context-budget controller. It cannot override source-first, compliance, citation, privacy, document-quality, or formal delivery gates.
+- Deep Reasoning Pass is an auditable decision summary, not private chain-of-thought and not a replacement for cognitive frameworks or self-review.
+- User-accepted skip/override records risk; it is not a quality pass.
+- This release does not include private project content, institution-specific requirements, participant material, private local paths, credentials, runtime state, or generated private reports.
+
 ## v1.5.2 - DOCX Structure And Layout Guards - 2026-06-09
 
 Status: formal Word delivery reliability update.