agentplane 0.6.2 → 0.6.4

package/assets/AGENTS.md

@@ -19,7 +19,6 @@ Detailed procedures live in canonical modules from `## CANONICAL DOCS`.
 ## PROJECT
 
 - Repository type: user project initialized with `agentplane`.
-- Gateway role: keep this file compact and deterministic; move scenario-specific details to policy modules.
 - CLI rule: prefer `ap` for compact agent-oriented commands; fall back to `agentplane`; if neither is available, stop and request installation guidance (do not invent repo-local entrypoints).
 - Startup shortcut: run `## COMMANDS -> Preflight`, then use `ap quickstart`; activate `ap role ORCHESTRATOR` for planning and `ap role <ROLE>` for the active owner before owner-scoped execution; then apply `## LOAD RULES` before any mutation. The guarded route is determined by `workflow.mode` in `.agentplane/WORKFLOW.md`; use `ap quickstart` as the canonical summary of the active path before mutating. In `branch_pr`, start from `ap work start ... --worktree`; in `direct`, stay in the current checkout and use the task lifecycle route.
 
@@ -85,6 +84,7 @@ git commit -m "Implement <task>"
 ap task verify-show <task-id>
 ap pr open <task-id> --branch task/<task-id>/<slug> --author <ROLE>
 ap verify <task-id> --ok|--rework --by <ROLE> --note "..."
+ap verify <task-id> --ok|--rework --by EVALUATOR --note "..." # verify --by EVALUATOR
 ap integrate <task-id> --branch task/<task-id>/<slug> --run-verify
 ap finish <task-id> --author INTEGRATOR --body "Verified: ..." --result "..." --commit <git-rev> --close-commit
 ```
@@ -145,7 +145,8 @@ Condition: task includes mutation (file edits, task-state changes, commits, merg
 5. IF task modifies implementation code paths THEN LOAD `@.agentplane/policy/dod.code.md`.
 6. IF task modifies docs/policy-only paths (`AGENTS.md`, docs, `.agentplane/policy/**`) THEN LOAD `@.agentplane/policy/dod.docs.md`.
 7. IF task modifies policy files (`AGENTS.md` or `.agentplane/policy/**`) THEN LOAD `@.agentplane/policy/governance.md`.
-8. IF task modifies `.agentplane/policy/incidents.md` THEN LOAD `@.agentplane/policy/incidents.md`.
+8. IF task reads, writes, learns, curates, verifies, or relies on local context THEN LOAD `@.agentplane/policy/context.must.md`.
+9. IF task modifies `.agentplane/policy/incidents.md` THEN LOAD `@.agentplane/policy/incidents.md`.
 
 Routing constraints:
 
@@ -219,14 +220,13 @@ Detailed DoD rules are in `.agentplane/policy/dod.core.md`, `.agentplane/policy/
 - DOC `.agentplane/policy/workflow.release.md`
 - DOC `.agentplane/policy/workflow.upgrade.md`
 - DOC `.agentplane/policy/security.must.md`
+- DOC `.agentplane/policy/context.must.md`
 - DOC `.agentplane/policy/dod.core.md`
 - DOC `.agentplane/policy/dod.code.md`
 - DOC `.agentplane/policy/dod.docs.md`
 - DOC `.agentplane/policy/governance.md`
 - DOC `.agentplane/policy/incidents.md`
 
----
-
 <!-- /ap:fragment -->
 <!-- ap:fragment id="gateway.agents.example.reference.examples" slot="example" mutability="replaceable" -->
 

package/assets/agents/CURATOR.json

@@ -22,7 +22,7 @@
   "workflow": {
     "goal": "Goal: assimilate approved raw sources into readable wiki memory, durable sourced derived knowledge, and reusable incident advice without treating search indexes, raw harvest scaffolds, or caches as truth.",
     "success.criteria": "Success criteria: selected source files and policies are read; existing wiki/facts/graph/capabilities/incidents are searched before new artifacts are written; every factual wiki/derived/incident claim has source_refs or an explicit no-source reason; useful wiki cross-links are added where existing target pages or glossary anchors improve navigation; contradictions and open questions are recorded; changed paths stay inside allowed_outputs and outside forbidden_outputs; context verify-task, context doctor, graph validation, incident validation, and ACR checks are run or concrete skips are recorded.",
-    "constraints": "Constraints: mutate knowledge only through the active AgentPlane task; never edit context/raw by default; never write .agentplane/context/service; preserve the chosen wiki hierarchy unless new evidence makes it misleading; keep glossary entries as a thin index over wiki pages and graph entities, not as competing truth; do not silently delete or merge facts/entities/edges/incidents; do not copy private raw content into README, ACR, reports, wiki, or incident advice unless policy permits redacted output; propose capabilities only when frontmatter allows it; let deterministic code validate registry shape, dedupe, required tags, and allowed commands.",
+    "constraints": "Constraints: mutate knowledge only through the active AgentPlane task; never edit context/raw by default; never write .agentplane/context/service; preserve the chosen wiki hierarchy unless new evidence makes it misleading; keep glossary entries as a thin index over wiki pages and graph entities, not as competing truth; do not silently delete or merge facts/entities/edges/incidents; do not copy secrets or non-publishable source spans into README, ACR, reports, wiki, or incident advice; propose capabilities only when frontmatter allows it; let deterministic code validate registry shape, dedupe, required tags, and allowed commands.",
     "stop.rules": "Stop rules: stop on missing source refs, unclear redaction policy, forbidden output requirements, stale or unresolvable source refs, capability or incident promotion beyond local proposal, or verification failures that would make the knowledge layer unreviewable.",
     "output": "Output: sources read, wiki pages changed, facts/entities/edges/provenance rows changed, incident advice matches/promotions when any, capability proposals when any, verification commands with results, residual contradictions/open questions, and rollback notes."
   }

package/assets/agents/EVALUATOR.json

@@ -18,7 +18,7 @@
   },
   "workflow": {
     "goal": "Goal: decide whether the latest task or eval attempt satisfies the documented quality contract without relying on the runner's self-claim alone.",
-    "success.criteria": "Success criteria: required task sections and Verify Steps are mapped to concrete evidence; result manifest and artifacts are structurally valid; hard policy/security/lifecycle gates pass; LLM quality scoring is used only where the approved rubric asks for judgement; the final verdict is reproducible from cited evidence.",
+    "success.criteria": "Success criteria: required task sections and Verify Steps are mapped to concrete evidence; result manifest and artifacts are structurally valid; hard policy/security/lifecycle gates pass; LLM quality scoring is used only where the approved rubric asks for judgement; context.maximum_assimilation work is checked for source-shaped wiki topology, useful Obsidian-compatible wikilinks, page granularity, line-addressed provenance, coverage gaps, glossary alias safety, raw-deletion resilience, and leakage risk; the final verdict is reproducible from cited evidence.",
     "constraints": "Constraints: use loaded gateway and policy modules as binding constraints; separate deterministic gates from LLM judgement; do not edit implementation files; do not finish or integrate tasks unless the approved plan explicitly assigns evaluator closure; preserve raw trace/artifact paths instead of copying assistant prose into task docs.",
     "stop.rules": "Stop rules: mark blocked when evidence is missing, stale, unverifiable, policy-sensitive, or outside approved scope; mark rework when criteria are testable but unmet; require human approval before changing pass criteria, promotion thresholds, or security-sensitive interpretation.",
     "output": "Output: verdict, failed or satisfied criteria, evidence paths, LLM judgement summary when used, rework context for the next runner pass, and finish/promote recommendation."

package/assets/policy/context.must.md

@@ -0,0 +1,69 @@
+<!-- ap:fragment id="policy.context.must.body.context.must" slot="body" mutability="replaceable" -->
+
+# Context MUST Rules
+
+Use this module when a task reads, writes, learns, curates, verifies, or relies on AgentPlane local context.
+
+Local context means `context/wiki/**`, `context/capabilities/**`, `context/raw/**`, `.agentplane/context/derived/facts/**`, `.agentplane/context/derived/graph/**`, `.agentplane/context/derived/capabilities/**`, and `.agentplane/context/**`.
+
+<!-- /ap:fragment -->
+<!-- ap:fragment id="policy.context.must.hard_constraint.source.first" slot="hard_constraint" mutability="append_only" -->
+
+## Source-first contract
+
+- MUST search existing context before creating new wiki pages, facts, graph entities, capabilities, or aliases.
+- MUST prefer updating an existing canonical page/entity over creating duplicates.
+- MUST keep raw source material, synthesized wiki prose, structured facts/graph rows, and capability notes distinct.
+- MUST mark facts, decisions, risks, workflows, and definitions with source refs or an explicit no-source reason.
+- MUST preserve conflicts and open questions instead of flattening contradictory sources into one unsourced claim.
+
+<!-- /ap:fragment -->
+<!-- ap:fragment id="policy.context.must.commands.command.contract" slot="commands" mutability="replaceable" -->
+
+## Command contract
+
+```bash
+ap context search "<query>"
+ap context show <ref>
+ap context learn changes
+ap context learn files <path> [--run]
+ap context ingest --changed|--all|--index-only
+ap context reindex --include-raw
+ap context wiki lint <path>
+ap context wiki explain <path>
+ap context wiki link <path>
+ap context wiki index context/wiki
+ap context check
+ap context doctor
+ap context verify-task <task-id>
+```
+
+- Use `search` for discovery and `show` for exact source readback.
+- Use `learn` to create human-facing context assimilation tasks instead of ad hoc memory writes; use `ingest`/`reindex` for lower-level agent and automation pipelines.
+- Use wiki lint/explain/link/index when adding or changing wiki pages.
+- Run `verify-task` before closing tasks that create or rely on context artifacts.
+
+<!-- /ap:fragment -->
+<!-- ap:fragment id="policy.context.must.hard_constraint.write.boundaries" slot="hard_constraint" mutability="append_only" -->
+
+## Write boundaries
+
+- MUST NOT manually edit `.agentplane/context/derived/**`; rebuild projections through context commands.
+- MUST NOT write secrets or non-publishable source spans into public wiki pages, task summaries, ACRs, reports, or incident advice.
+- MUST preserve the user-created hierarchy under `context/raw/**` when citing raw sources.
+- MUST NOT create glossary-only truth; glossary entries are aliases/navigation over wiki pages and graph entities.
+- SHOULD use proposal-before-promotion for harvested or inferred context when provenance, staleness, or conflicts are not clean.
+
+<!-- /ap:fragment -->
+<!-- ap:fragment id="policy.context.must.check.dod" slot="check" mutability="append_only" -->
+
+## Context DoD
+
+Context work is done only when:
+
+1. Existing context was searched or a concrete skip reason was recorded.
+2. New or changed factual claims have source refs or explicit no-source reasons.
+3. Wiki links/indexes were refreshed when pages were added, moved, or materially renamed.
+4. Derived projection checks were run or skipped with reason and risk.
+5. `ap context verify-task <task-id>` passed for task-bound context work.
+<!-- /ap:fragment -->

package/dist/.build-manifest.json

@@ -2,7 +2,7 @@
   "schema_version": 1,
   "manifest_kind": "package",
   "package_name": "agentplane",
-  "package_version": "0.6.2",
-  "git_head": "38d20e02646371b587820adb328039c615323f47",
-  "watched_runtime_snapshot_hash": "d9122cf6ac941123139d675eac929456b8e4caf0ce3e811bc69af6ea9905ec2a"
+  "package_version": "0.6.4",
+  "git_head": "75072680dde8840f26ab3ae1fd02fae8b7e54bcd",
+  "watched_runtime_snapshot_hash": "2f2d00d089cd72157ee5cf45afa084df1ab2d65fe3b8fb3823b35d388b0bd2a6"
 }