@veewo/claw-core 0.2.10 → 0.2.12

package/dist/src/project-config-defaults.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"project-config-defaults.js","sourceRoot":"","sources":["../../src/project-config-defaults.ts"],"names":[],"mappings":"AAAA,MAAM,CAAC,MAAM,yBAAyB,GAAG,CAAC,CAAC;AAE3C,MAAM,UAAU,sBAAsB,CAAC,aAInB;IAClB,OAAO,aAAa,EAAE,MAAM,EAAE,OAAO,KAAK,KAAK,CAAC;AAClD,CAAC"}

package/dist/src/resources/cindy-delegate-writer/TEMPLATE.json

@@ -0,0 +1,49 @@
+{
+  "id": "internal-cindy-knowledge-delegate",
+  "version": "0.2.12",
+  "scope": "session",
+  "title": "internal Cindy knowledge delegate",
+  "status": "process.active",
+  "goal": {
+    "text": "Complete one dispatched Cindy knowledge finalization through its generated assignment subplan."
+  },
+  "requirements": {
+    "summary": "Claim the durable Cindy job, execute its frozen assignment template as a subplan, and persist one terminal result.",
+    "openQuestions": [],
+    "acceptanceCriteria": [
+      "The claim captures the originating Cindy task conclusions before issuing ownership.",
+      "The generated assignment template executes inside the session workflow.",
+      "The claimed job receives one matching terminal result."
+    ]
+  },
+  "tasks": [
+    {
+      "id": 1,
+      "title": "Claim the ready Cindy knowledge job",
+      "detail": "Use the claw-kit Ghost knowledge.claim operation with the exact finalization id supplied by dispatch. Retain claimToken, templatePath, assignments, and immutable configuration. If the originating report has not yet persisted, retry the same claim; if claimed is false, do not edit project knowledge.",
+      "status": "pending"
+    },
+    {
+      "id": 2,
+      "title": "Execute the assignment subplan",
+      "detail": "Use the claw-kit Ghost subplan.create operation with this finalizer task name as parent, task id 2, and the templateFile returned by knowledge.claim.",
+      "status": "pending"
+    },
+    {
+      "id": 3,
+      "title": "Persist the terminal result",
+      "detail": "After a successful claim, use the claw-kit Ghost knowledge.done operation once with the exact finalization id and claimToken. Report succeeded only when the assignment subplan succeeded; otherwise report failed with the remaining error.",
+      "status": "pending"
+    }
+  ],
+  "references": [],
+  "rules": [
+    "Use only the supplied finalization id, claim token, and generated assignment template.",
+    "Do not modify the supplied plan or report materials.",
+    "Do not start another finalizer or delegate the assignments to another agent."
+  ],
+  "keyDecisions": [],
+  "retrospective": {
+    "summary": ""
+  }
+}

package/dist/src/resources/delegate-writer/TEMPLATE.json

@@ -1,6 +1,6 @@
 {
   "id": "internal-knowledge-delegate",
-  "version": "0.2.10",
+  "version": "0.2.12",
   "scope": "session",
   "title": "internal knowledge delegate",
   "status": "process.active",

package/dist/src/resources/doc-updater/SKILL.md

@@ -0,0 +1,17 @@
+---
+name: doc-updater
+description: Update configured existing external documentation as the dependent documentation stage of knowledge finalization.
+---
+
+# Doc updater
+
+Resolve `<skill-dir>` as the directory containing this file.
+
+Use this skill only when it owns the external-document stage of an active
+knowledge-finalization plan. Create its subplan with:
+
+`claw subplan create --parent <parent-task-name> --task-id <id> --template-file "<skill-dir>/TEMPLATE.json"`
+
+The parent stage supplies the frozen external documentation paths, finalization
+materials, and the resulting canonical knowledge state. This skill has no
+standalone plan route.

package/dist/src/resources/doc-updater/TEMPLATE.json

@@ -0,0 +1,98 @@
+{
+  "id": "doc-updater",
+  "version": "0.2.12",
+  "scope": "session",
+  "title": "doc-updater",
+  "status": "process.active",
+  "goal": {
+    "text": "Update affected existing external documentation from the finalized project knowledge state."
+  },
+  "requirements": {
+    "summary": "Use the supplied finalization evidence and resulting canonical knowledge to repair stale or conflicting claims in the frozen external documentation corpus without creating documents or imposing a canonical format.",
+    "openQuestions": [],
+    "acceptanceCriteria": [
+      "Only existing documents inside the frozen external documentation paths are eligible for modification.",
+      "Current-state claims are checked against current implementation and resulting canonical knowledge without rewriting requirements, future design, or history as completed behavior.",
+      "Affected documents retain their own language, structure, purpose, and unrelated user edits.",
+      "Focused semantic and exhaustive identifier searches find no unresolved stale or competing current claim in the affected external documentation corpus."
+    ]
+  },
+  "tasks": [
+    {
+      "id": 1,
+      "title": "Identify affected existing documentation",
+      "detail": "Read this subplan's parent assignment detail to recover the supplied finalization materials and frozen external documentation paths, then inspect the resulting canonical knowledge state. Enumerate only existing documents inside those paths. Use claw search to find documents affected by the completed conclusions, then use exact identifiers, config keys, commands, paths, and other distinguishing terms to find plausible matches that top-k recall could omit. Do not modify documents yet.",
+      "status": "pending",
+      "guidance": {
+        "onDone": {
+          "default": {
+            "mergeMode": "override",
+            "summary": "Affected existing documents and plausible competing claims are identified; resolve their semantics and freshness next.",
+            "nextsteps": [
+              "Carry forward the evidence location and affected passage for each candidate.",
+              "Treat absence of an existing affected document as a valid no-edit outcome rather than creating a new document."
+            ],
+            "nextTaskId": 2
+          }
+        }
+      }
+    },
+    {
+      "id": 2,
+      "title": "Resolve document semantics and freshness",
+      "detail": "Classify each affected passage as current-state documentation, historical evidence, a requirement, future design, tutorial guidance, or an example. Check material current-state claims against current implementation anchors, relevant working-tree changes, and the resulting canonical Truth and ADR state. Decide whether each passage must be updated, narrowed, time-bound, or retained unchanged. Do not mechanically overwrite an intentional requirement, future design, or useful history because its wording differs from current implementation.",
+      "status": "pending",
+      "guidance": {
+        "onDone": {
+          "default": {
+            "mergeMode": "override",
+            "summary": "Affected passages are classified and freshness-qualified; update only the safely resolved existing documents.",
+            "nextsteps": [
+              "Preserve each document's existing purpose and local conventions.",
+              "Retain unresolved ambiguity as an explicit no-edit reason instead of guessing."
+            ],
+            "nextTaskId": 3
+          }
+        }
+      }
+    },
+    {
+      "id": 3,
+      "title": "Update existing external documents",
+      "detail": "Apply the freshness-qualified changes to existing eligible documents only. Preserve their language, structure, formatting, scope, and unrelated user edits. Do not create, move, rename, or standardize documents, and do not require owner metadata or canonical Truth/ADR formatting. Keep useful historical or future-facing content when it is accurately scoped.",
+      "status": "pending",
+      "guidance": {
+        "onDone": {
+          "default": {
+            "mergeMode": "override",
+            "summary": "Eligible existing documents are updated without changing corpus shape; verify the affected corpus now.",
+            "nextsteps": [
+              "Re-run both semantic and exhaustive identifier searches.",
+              "Check that edited current-state wording agrees with implementation and canonical knowledge."
+            ],
+            "nextTaskId": 4
+          }
+        }
+      }
+    },
+    {
+      "id": 4,
+      "title": "Verify external documentation consistency",
+      "detail": "Re-run focused claw search and exhaustive identifier search across the frozen external documentation paths. Verify that affected current-state claims no longer conflict with current implementation or resulting canonical knowledge, that historical and future-facing content remains correctly scoped, and that no file outside the frozen paths was modified. Return changed paths, evidence-backed no-edit reasons, and any conflict that could not be resolved safely.",
+      "status": "pending"
+    }
+  ],
+  "references": [],
+  "rules": [
+    "Modify only existing documents inside the frozen external documentation paths; never create, move, or rename a document.",
+    "Do not assume an owner, frontmatter schema, section layout, language, or canonical document format.",
+    "Distinguish current state from requirements, future design, examples, and history before editing.",
+    "Current implementation and resulting canonical knowledge qualify current-state claims, but neither justifies erasing accurately scoped intent or history.",
+    "Preserve unrelated user edits and local document conventions.",
+    "Do not report an unresolved material current-state conflict as governed successfully."
+  ],
+  "keyDecisions": [],
+  "retrospective": {
+    "summary": ""
+  }
+}

package/dist/src/resources/knowledge-writer/CONTENT-COVERAGE.md

@@ -1,30 +1,30 @@
-# knowledge-writer content coverage
-
-## Source to converted-home mapping
-
-- Trigger and template entry: `SKILL.md`; input and quality contract: `TEMPLATE.json` requirements, tasks, guidance, and rules.
-- Template compatibility: `TEMPLATE.json` declares the current claw CLI version.
-- Conclusion evidence from every supplied material, with task-status interpretation when present: `non-claw-fallback.md`, template tasks 1 and 2, acceptance criteria, and template rules.
-- Evidence freshness and unresolved-conflict handling: `non-claw-fallback.md`, template task 2, acceptance criteria, and template rules.
-- Runtime scope: `TEMPLATE.json` top-level `scope` only; it is not repeated in writer prompts.
-- Fixed Truth-then-ADR sequence without routing choices: template tasks 4 and 5 and `non-claw-fallback.md`.
-- Machine-stable current/history/superseded metadata and dated evolution grammar: `knowledge-format.md`, template tasks 4 and 5, and `non-claw-fallback.md`.
-- On-write format inspection and repair without corpus-wide migration: `knowledge-format.md`, template tasks 3 through 6, acceptance criteria, and template rules.
-- Canonical-owner discovery and exhaustive search: template tasks 3 through 6 plus template rules.
-- One-owner stewardship and cross-document consistency: template task 6, template rules, and `non-claw-fallback.md`.
-- Writing constraints, encoding, exact identifiers, and return behavior: `non-claw-fallback.md`, template task 6, and template rules.
-- Supplied-material immutability: template task 1, acceptance criteria, template rules, and `non-claw-fallback.md`.
-- Verification gate: template task 6 requires focused and exhaustive post-write review before completion.
-
-## Coverage result
-
-- [x] Important source triggers and inputs are represented.
-- [x] Conclusion-bearing content from every supplied material drives deposition; task status informs interpretation without turning task metadata into an execution record.
-- [x] The ordered workflow always performs the Truth pass before the ADR pass and has no route-choice task.
-- [x] Search tools, ownership constraints, and safety boundaries are represented.
-- [x] Time-bounded authority, current-anchor checks, and unresolved-freshness no-edit behavior are represented.
-- [x] Writing and consistency verification requirements are represented.
-- [x] Canonical state and dated evolution semantics are represented.
-- [x] Every written owner is repaired to canonical format while untouched documents remain unmigrated.
-- [x] Full direct behavior remains available in the adjacent fallback.
-- [x] No source companion scripts or external links require migration.
+# knowledge-writer content coverage
+
+## Source to converted-home mapping
+
+- Trigger and template entry: `SKILL.md`; input and quality contract: `TEMPLATE.json` requirements, tasks, guidance, and rules.
+- Template compatibility: `TEMPLATE.json` declares the current claw CLI version.
+- Conclusion evidence from every supplied material, with task-status interpretation when present: `non-claw-fallback.md`, template tasks 1 and 2, acceptance criteria, and template rules.
+- Evidence freshness and unresolved-conflict handling: `non-claw-fallback.md`, template task 2, acceptance criteria, and template rules.
+- Runtime scope: `TEMPLATE.json` top-level `scope` only; it is not repeated in writer prompts.
+- Fixed Truth-then-ADR sequence without routing choices: template tasks 4 and 5 and `non-claw-fallback.md`.
+- Machine-stable current/history/superseded metadata and dated evolution grammar: `knowledge-format.md`, template tasks 4 and 5, and `non-claw-fallback.md`.
+- On-write format inspection and repair without corpus-wide migration: `knowledge-format.md`, template tasks 3 through 6, acceptance criteria, and template rules.
+- Canonical-owner discovery and exhaustive search: template tasks 3 through 6 plus template rules.
+- One-owner stewardship and cross-document consistency: template task 6, template rules, and `non-claw-fallback.md`.
+- Writing constraints, encoding, exact identifiers, and return behavior: `non-claw-fallback.md`, template task 6, and template rules.
+- Supplied-material immutability: template task 1, acceptance criteria, template rules, and `non-claw-fallback.md`.
+- Verification gate: template task 6 requires focused and exhaustive post-write review before completion.
+
+## Coverage result
+
+- [x] Important source triggers and inputs are represented.
+- [x] Conclusion-bearing content from every supplied material drives deposition; task status informs interpretation without turning task metadata into an execution record.
+- [x] The ordered workflow always performs the Truth pass before the ADR pass and has no route-choice task.
+- [x] Search tools, ownership constraints, and safety boundaries are represented.
+- [x] Time-bounded authority, current-anchor checks, and unresolved-freshness no-edit behavior are represented.
+- [x] Writing and consistency verification requirements are represented.
+- [x] Canonical state and dated evolution semantics are represented.
+- [x] Every written owner is repaired to canonical format while untouched documents remain unmigrated.
+- [x] Full direct behavior remains available in the adjacent fallback.
+- [x] No source companion scripts or external links require migration.

package/dist/src/resources/knowledge-writer/SKILL.md

@@ -1,9 +1,12 @@
----
-name: knowledge-writer
-description: Evaluate supplied materials by their content, then maintain canonical Truth followed by ADR knowledge in one consistency-aware pass. Use only when explicitly invoked with supplied materials; do not trigger this skill implicitly.
----
-# Knowledge writer
-
-Resolve `<skill-dir>` as the directory containing this file. Run `claw plan create --template-file "<skill-dir>/TEMPLATE.json" --title "knowledge-writer"`, then follow the returned `workflowGuidance`.
-
-If the template or claw CLI is unavailable, follow `non-claw-fallback.md` directly.
+---
+name: knowledge-writer
+description: Evaluate supplied materials by their content, then maintain canonical Truth followed by ADR knowledge in one consistency-aware pass. Use only when explicitly invoked with supplied materials; do not trigger this skill implicitly.
+---
+# Knowledge writer
+
+Resolve `<skill-dir>` as the directory containing this file.
+
+- When this skill owns the supplied finalization assignment as a stage of an active plan, run `claw subplan create --parent <parent-task-name> --task-id <id> --template-file "<skill-dir>/TEMPLATE.json"`.
+- When explicitly invoked with supplied materials outside an active parent plan, run `claw plan create --template-file "<skill-dir>/TEMPLATE.json" --title "knowledge-writer"`.
+
+If the template or claw CLI is unavailable, follow `non-claw-fallback.md` directly.

package/dist/src/resources/knowledge-writer/TEMPLATE.json

@@ -1,6 +1,6 @@
 {
   "id": "knowledge-writer",
-  "version": "0.2.10",
+  "version": "0.2.12",
   "scope": "session",
   "title": "knowledge-writer",
   "status": "process.active",

package/dist/src/resources/knowledge-writer/knowledge-format.md

@@ -1,47 +1,95 @@
-# Canonical knowledge evolution format
-
-Use this format for every new Truth or ADR document and every existing owner written by the current writer pass. Inspect the selected owner before writing and repair nonconforming structure in the same edit. Do not mass-rewrite untouched documents.
-
-## Document metadata
-
-Document kind is inferred from its canonical path: documents under `adr/` are ADRs; other documents under the Truth root are Truth documents. Do not add a document-kind field.
-
-Truth documents default to `current`. ADR documents default to `accepted`. When a different document state must be explicit, place a renderer-hidden generic comment immediately after the title:
-
-```markdown
-# ADR: Replaced decision
-
-<!-- document-state: superseded -->
-```
-
-Allowed document states are:
-
-- Truth: `current` or `historical`
-- ADR: `accepted`, `superseded`, or `historical`
-
-An ordinary leading `## Status` or `## 状态` section may also supply the document state. Do not emit tool-specific metadata names.
-
-## Current and evolution sections
-
-Use renderer-hidden, machine-stable ASCII comments before natural-language headings:
-
-```markdown
-<!-- state: current -->
-## Current behavior
-
-Current canonical content.
-
-<!-- state: history -->
-## Evolution history
-
-<!-- dated: 2026-07-20 -->
-### Replaced the earlier route
-
-The prior behavior and why it matters for rollback or repeated work.
-```
-
-For ADRs, the ordinary Context, Decision, Alternatives, and Consequences sections inherit the document state. Mark only retained decision evolution with `<!-- state: history -->`.
-
-Each `<!-- dated: YYYY-MM-DD -->` comment starts one complete evolution unit at the level-three heading that follows it. The unit extends to the next dated comment or the next level-two heading. The date is a stable checkpoint label and search signal, not a time-to-live value. Multiple checkpoints may use the same date. Their canonical evolution order is document order.
-
-Do not append an evolution unit for routine progress or every successful turn. Add one only when the former fact or decision remains useful for rollback, compatibility, feature repetition, incident reasoning, or understanding a meaningful transition.
+# Canonical Truth and ADR format
+
+Use this format for every new Truth or ADR document and every existing owner
+written by the current governance pass. Repair a selected owner's format in the
+same edit, but do not mass-rewrite untouched documents.
+
+## Kind, path, and state
+
+Document kind is inferred from its canonical path: documents under `adr/` are
+ADRs; other documents under the Truth root are Truth documents. Do not add a
+document-kind field.
+
+Truth defaults to `current`. ADR defaults to `accepted`. When another state is
+required, place a renderer-hidden comment immediately after the title:
+
+```markdown
+# ADR: Replaced decision
+
+<!-- document-state: superseded -->
+```
+
+Allowed states:
+
+- Truth: `current`, `historical`
+- ADR: `accepted`, `superseded`, `historical`
+
+An ordinary leading `## Status` or `## 状态` section may also supply the state.
+Do not emit tool-specific metadata names.
+
+## Truth structure
+
+```markdown
+# Feature or behavior title
+
+<!-- state: current -->
+## Current behavior
+
+Current canonical facts, constraints, ownership, implementation anchors, and
+verification rules.
+
+<!-- state: history -->
+## Evolution history
+
+<!-- dated: 2026-08-05 -->
+### Replaced behavior
+
+The former fact and why it remains useful.
+```
+
+Keep one current owner for each material fact. Move former behavior into
+history only when it remains useful for rollback, compatibility, repeated
+work, incident reasoning, or understanding a meaningful transition.
+
+## ADR structure
+
+```markdown
+# ADR: Decision title
+
+## Context
+
+The forces and problem that required a durable decision.
+
+## Decision
+
+The accepted decision and its ownership boundary.
+
+## Alternatives
+
+The material alternatives and why they were not selected.
+
+## Consequences
+
+The tradeoffs, follow-up constraints, and verification implications.
+
+<!-- state: history -->
+## Decision evolution
+
+<!-- dated: 2026-08-05 -->
+### Earlier decision
+
+The superseded choice and why retaining it is useful.
+```
+
+Context, Decision, Alternatives, and Consequences inherit the document state.
+Do not manufacture an ADR when no durable decision exists.
+
+## Evolution grammar
+
+Place `<!-- state: current -->` and `<!-- state: history -->` before natural
+language headings. Each `<!-- dated: YYYY-MM-DD -->` comment starts one complete
+evolution unit at the following level-three heading. The unit ends at the next
+dated comment or level-two heading. The date is a stable checkpoint and search
+signal, not a time-to-live value; multiple checkpoints may share a date.
+
+Do not append history for routine progress or every successful task.

package/dist/src/resources/knowledge-writer/non-claw-fallback.md

@@ -1,53 +1,53 @@
-# Knowledge writer fallback
-
-Act as the knowledge-base steward for the project. Leave the relevant Truth and ADR corpus accurate, coherent, and easy for future agents to trust.
-
-## Input and evidence boundary
-
-Read every supplied material completely. Interpret evidence from its content and semantics rather than requiring a particular filename, field, record shape, or serialization format. Extract executed conclusions, verified findings, retrospective lessons, key decisions, and other explicit outcomes wherever they appear. Do not repeat implementation or test verification or modify supplied materials.
-
-When task status is present, it helps interpret completed, pending, and blocked scope, but a task list is not an execution log. Infer execution results, verified findings, planning outcomes, and durable decisions from conclusion-bearing content; never turn task titles, descriptions, requirements, or intentions into completed results merely because they appear in an input.
-
-## Evidence freshness
-
-Trusted means the evidence was verified at the revision or worktree state it describes; it does not make an older report permanently authoritative for current behavior.
-
-- Before writing current behavior from a supplied conclusion, read the relevant implementation anchors and inspect later or overlapping working-tree changes. This is a freshness check, not repeated implementation verification; do not rerun tests merely to reconfirm the report.
-- When current implementation supersedes the report, current implementation owns current-state Truth and the report may only support historical or version-bound evidence.
-- If chronology or authority cannot be resolved safely, omit the affected canonical write and retain the freshness conflict as the reason.
-
-## Fixed deposition sequence
-
-Process the eligible evidence without a route choice:
-
-1. Maintain Truth for stable behavior, architecture facts, constraints, pitfalls, code anchors, and verification rules.
-2. Then maintain ADRs for durable decisions, context, rationale, alternatives, tradeoffs, ownership, and consequences, using the same evidence and resulting Truth state.
-3. Finally review Truth and ADR together for ownership and consistency.
-
-Use `knowledge-format.md` for every new document and every existing owner written by this pass. Inspect each selected owner before writing and repair nonconforming structure in the same edit; leave untouched documents unmigrated. Dated identifies an evolution checkpoint, not age or time-to-live.
-
-It is valid for either pass to make no edit when the eligible evidence contains no new or changed durable knowledge. Temporary progress, speculation, conversational narration, unchanged facts, and unfinished-task claims do not belong in canonical knowledge.
-
-## Stewardship and ownership
-
-Use `claw search` to discover existing owners, then open every plausible candidate before judging it. Use exhaustive text search for distinguishing identifiers when top-k recall could hide another current claim. Update the document that already owns the topic; create a new document only for a genuinely new durable topic after filename and title collision checks.
-
-Maintain one current owner for each material fact or decision. A broad or neighboring document is still a competing owner when it restates an unqualified current rule. Reconcile overlaps by extending the canonical owner and narrowing other current claims to references, historical evidence, or non-overlapping scope. Preserve unrelated user edits and repository conventions.
-
-Truth and ADR are one knowledge system: after both ordered passes, review the related documents together and resolve contradictions between current-state Truth, decision ownership, and consequences. Do not report completion while a material current claim remains inconsistent or ownerless.
-
-## Writing and verification
-
-- Follow the repository's language and document shape.
-- Preserve exact identifiers, config keys, commands, and error text.
-- Use project-relative paths in canonical documents.
-- Ground every fact, path, owner, and alternative in supplied or inspected evidence.
-- Label historical and superseded evidence explicitly; never silently promote it to current behavior.
-- Add dated evolution only for former facts or decisions that remain useful for rollback, compatibility, feature repetition, incident reasoning, or understanding a meaningful transition.
-- Repair every written Truth or ADR owner to the canonical format in `knowledge-format.md`; do not defer a discovered format mismatch.
-- Repair mojibake and preserve valid Markdown encoding.
-- Re-run focused and exhaustive searches after writing; every plausible hit must be the selected owner, an explicit reference, historical or version-bound evidence, or a narrowed non-overlapping claim.
-
-## Return
-
-Return a brief completion note with changed paths or the evidence-backed no-edit reasons. Response format is not part of the contract.
+# Knowledge writer fallback
+
+Act as the knowledge-base steward for the project. Leave the relevant Truth and ADR corpus accurate, coherent, and easy for future agents to trust.
+
+## Input and evidence boundary
+
+Read every supplied material completely. Interpret evidence from its content and semantics rather than requiring a particular filename, field, record shape, or serialization format. Extract executed conclusions, verified findings, retrospective lessons, key decisions, and other explicit outcomes wherever they appear. Do not repeat implementation or test verification or modify supplied materials.
+
+When task status is present, it helps interpret completed, pending, and blocked scope, but a task list is not an execution log. Infer execution results, verified findings, planning outcomes, and durable decisions from conclusion-bearing content; never turn task titles, descriptions, requirements, or intentions into completed results merely because they appear in an input.
+
+## Evidence freshness
+
+Trusted means the evidence was verified at the revision or worktree state it describes; it does not make an older report permanently authoritative for current behavior.
+
+- Before writing current behavior from a supplied conclusion, read the relevant implementation anchors and inspect later or overlapping working-tree changes. This is a freshness check, not repeated implementation verification; do not rerun tests merely to reconfirm the report.
+- When current implementation supersedes the report, current implementation owns current-state Truth and the report may only support historical or version-bound evidence.
+- If chronology or authority cannot be resolved safely, omit the affected canonical write and retain the freshness conflict as the reason.
+
+## Fixed deposition sequence
+
+Process the eligible evidence without a route choice:
+
+1. Maintain Truth for stable behavior, architecture facts, constraints, pitfalls, code anchors, and verification rules.
+2. Then maintain ADRs for durable decisions, context, rationale, alternatives, tradeoffs, ownership, and consequences, using the same evidence and resulting Truth state.
+3. Finally review Truth and ADR together for ownership and consistency.
+
+Use `knowledge-format.md` for every new document and every existing owner written by this pass. Inspect each selected owner before writing and repair nonconforming structure in the same edit; leave untouched documents unmigrated. Dated identifies an evolution checkpoint, not age or time-to-live.
+
+It is valid for either pass to make no edit when the eligible evidence contains no new or changed durable knowledge. Temporary progress, speculation, conversational narration, unchanged facts, and unfinished-task claims do not belong in canonical knowledge.
+
+## Stewardship and ownership
+
+Use `claw search` to discover existing owners, then open every plausible candidate before judging it. Use exhaustive text search for distinguishing identifiers when top-k recall could hide another current claim. Update the document that already owns the topic; create a new document only for a genuinely new durable topic after filename and title collision checks.
+
+Maintain one current owner for each material fact or decision. A broad or neighboring document is still a competing owner when it restates an unqualified current rule. Reconcile overlaps by extending the canonical owner and narrowing other current claims to references, historical evidence, or non-overlapping scope. Preserve unrelated user edits and repository conventions.
+
+Truth and ADR are one knowledge system: after both ordered passes, review the related documents together and resolve contradictions between current-state Truth, decision ownership, and consequences. Do not report completion while a material current claim remains inconsistent or ownerless.
+
+## Writing and verification
+
+- Follow the repository's language and document shape.
+- Preserve exact identifiers, config keys, commands, and error text.
+- Use project-relative paths in canonical documents.
+- Ground every fact, path, owner, and alternative in supplied or inspected evidence.
+- Label historical and superseded evidence explicitly; never silently promote it to current behavior.
+- Add dated evolution only for former facts or decisions that remain useful for rollback, compatibility, feature repetition, incident reasoning, or understanding a meaningful transition.
+- Repair every written Truth or ADR owner to the canonical format in `knowledge-format.md`; do not defer a discovered format mismatch.
+- Repair mojibake and preserve valid Markdown encoding.
+- Re-run focused and exhaustive searches after writing; every plausible hit must be the selected owner, an explicit reference, historical or version-bound evidence, or a narrowed non-overlapping claim.
+
+## Return
+
+Return a brief completion note with changed paths or the evidence-backed no-edit reasons. Response format is not part of the contract.

package/dist/src/templates/plans/default.js

@@ -1,6 +1,6 @@
 export const defaultPlanTemplate = {
     id: "default",
-    version: "0.2.10",
+    version: "0.2.12",
     status: "process.discussing",
     goal: {
         text: "",

package/dist/src/types.d.ts

@@ -36,6 +36,7 @@ export type ProjectConfig = {
     contextPaths?: string[];
     memory?: {
         enabled?: boolean;
+        autoUpdate?: boolean;
         externalDocPaths?: string[];
         embedding?: MemoryEmbeddingConfig | null;
     };