chore: standardize verdict field and add exploration rules across agents

.github/plugin/marketplace.json

@@ -1255,7 +1255,7 @@
       "name": "gem-team",
       "source": "plugins/gem-team",
       "description": "Self-Learning Multi-agent orchestration framework for spec-driven development and automated verification. With smarter tool calling and leaner context.",
-      "version": "1.138.0"
+      "version": "1.139.0"
     },
     {
       "name": "gesture-review",

agents/gem-browser-tester.agent.md

@@ -21,7 +21,7 @@ No improvisation.
 - Per scenario: navigate (pre-flight on first), precondition, fixture, flow (observe->act->verify), assert state/DB/API/visual reg.
 - On failure: capture screenshots, traces, logs. On success: retain/compare baselines. Store only if `evidence_required` is true.
 - Per page finalize: console errors, network failures, a11y audit (cache by semantic DOM hash). Only run `checks_to_run`.
-- Cleanup: close contexts, remove orphans, stop traces, persist evidence.
+- Cleanup: close contexts, remove orphans, finalize traces, persist finalized evidence (including failure traces).
 - Output: raw JSON per `output_format`. No markdown, no prose.
 </workflow>
 
@@ -37,7 +37,9 @@ No improvisation.
   "a11y_issues": 0,
   "handoff": {
     "evidence_path?": "string",
-    "verdict": "pass | fail | skip"
+    "verdict": "pass | fail | blocked",
+    "relevant_context": ["string"],
+    "stable_findings": [{ "finding": "string", "confidence": 0.95, "stable": true }]
   },
   "learn": "string"
 }
@@ -57,5 +59,6 @@ No improvisation.
 - `evidence_path?` is optional: emit only when `evidence_required` is true and artifacts were produced; omit otherwise.
 - Emit one-line `learn` on new failure mode, repeated blocker, or confirmed architecture fact; otherwise omit.
 - If a check is explicitly required but cannot run, report as blocker - never skip silently.
+- Explore only what the handoff does not cover. If a field is empty or missing, treat it as a gap.
 - Check relevant memory when applicable; expand as warranted.
 </rules>

agents/gem-code-simplifier.agent.md

@@ -38,11 +38,13 @@ No improvisation.
 {
   "status": "completed | failed | needs_retry | blocked",
   "reason": "string",
-  "fail": "fixable | needs_replan | escalate | flaky | regression | new_failure | platform_specific",
+  "fail": "fixable | needs_replan | escalate | flaky | regression | new_failure | platform_specific | test_bug",
   "handoff": {
     "changed_files": ["string"],
     "complexity_delta": 0,
-    "verdict": "pass | fail"
+    "verdict": "pass | fail | blocked",
+    "relevant_context": ["string"],
+    "stable_findings": [{ "finding": "string", "confidence": 0.95, "stable": true }]
   },
   "learn": "string"
 }
@@ -62,5 +64,7 @@ No improvisation.
 - Emit one-line `learn` on new failure mode, repeated blocker, or confirmed architecture fact; otherwise omit.
 - Prefer maintained official/in-stack libraries to custom code.
 - Fix code, not comment on it. Refactor only; add no features.
+- Explore only what the handoff does not cover. If a field is empty or missing, treat it as a gap.
+- If `handoff.global_findings` contains project conventions: apply them, don't rediscover.
 - Check relevant memory when applicable; expand as warranted.
 </rules>

agents/gem-debugger.agent.md

@@ -34,10 +34,10 @@ No improvisation.
     "debugger_diagnosis": {
       "root_cause": "string",
       "target_files": ["string"],
-      "reproduction": { "steps": ["string"], "expected": "string", "actual": "string" },
+      "reproduction?": { "steps": ["string"], "expected": "string", "actual": "string" },
       "fix_recommendations": ["string"]
     },
-    "lint_rule_recommendations": [{ "name": "string", "type": "built-in | custom", "files": ["string"] }]
+    "lint_rule_recommendations?": [{ "name": "string", "type": "built-in | custom", "files": ["string"] }]
   },
   "learn": "string"
 }
@@ -57,5 +57,7 @@ No improvisation.
 - Emit one-line `learn` on new failure mode, repeated blocker, or confirmed architecture fact; otherwise omit.
 - Stop when root cause reproduces in >=2 independent checks, or single definitive evidence (stack trace to root line) identifies it.
 - Investigate only when needed; every additional check must resolve an uncertainty, perform required work, or verify a result.
+- Explore only what the handoff does not cover. If a field is empty or missing, treat it as a gap.
+- If `handoff.global_findings` contains relevant architecture facts: use as context, don't re-trace.
 - Check relevant memory when applicable; expand as warranted.
 </rules>

agents/gem-devops.agent.md

@@ -31,10 +31,12 @@ No improvisation.
 {
   "status": "completed | failed | needs_retry | blocked",
   "reason": "string",
-  "fail": "fixable | needs_replan | escalate | flaky | regression | new_failure | platform_specific",
+  "fail": "fixable | needs_replan | escalate | flaky | regression | new_failure | platform_specific | test_bug",
   "handoff": {
     "evidence_path?": "string",
-    "verdict": "pass | fail | blocked"
+    "verdict": "pass | fail | blocked",
+    "relevant_context": ["string"],
+    "stable_findings": [{ "finding": "string", "confidence": 0.95, "stable": true }]
   },
   "learn": "string"
 }
@@ -54,5 +56,7 @@ No improvisation.
 - `evidence_path?` is optional: emit only when `evidence_required` is true and artifacts were produced; omit otherwise.
 - Emit one-line `learn` on new failure mode, repeated blocker, or confirmed architecture fact; otherwise omit.
 - Make operations idempotent, preferably atomic.
+- Explore only what the handoff does not cover. If a field is empty or missing, treat it as a gap.
+- If `handoff.plan_context.tech_stack` present: use for provider/workload classification.
 - Check relevant memory when applicable; expand as warranted.
 </rules>

agents/gem-documentation-writer.agent.md

@@ -31,13 +31,15 @@ Write docs, READMEs, API docs, diagrams. Maintain `AGENTS.md`. Never implement c
 {
   "status": "completed | failed | needs_retry | blocked",
   "reason": "string",
-  "fail": "fixable | needs_replan | escalate | flaky | regression | new_failure | platform_specific",
+  "fail": "fixable | needs_replan | escalate | flaky | regression | new_failure | platform_specific | test_bug",
   "created": 0,
   "updated": 0,
   "handoff": {
     "created_paths": ["string"],
     "updated_paths": ["string"],
-    "verdict": "pass | fail"
+    "verdict": "pass | fail | blocked",
+    "relevant_context": ["string"],
+    "stable_findings": [{ "finding": "string", "confidence": 0.95, "stable": true }]
   },
   "learn": "string"
 }
@@ -61,5 +63,7 @@ Write docs, READMEs, API docs, diagrams. Maintain `AGENTS.md`. Never implement c
 - No buzzwords ("AI Powered", "Revolutionary", "Seamless", etc.). Use specific language.
 - Every section must exist because the product needs it. Remove template filler.
 - No fabricated statistics or claims. Use `[REAL DATA]` or omit the claim.
+- Explore only what the handoff does not cover. If a field is empty or missing, treat it as a gap.
+- If `handoff.global_findings` contains API contracts: document those, don't re-extract.
 - Check relevant memory when applicable; expand as warranted.
 </rules>

agents/gem-implementer.agent.md

@@ -58,11 +58,14 @@ No improvisation.
 - Comments: justify non-obvious logic; include required lint directives and generated-file markers; don't restate what the code shows.
 - YAGNI/KISS: minimum that meets the task; reuse before building; justify every extra layer/agent/task/wave barrier; never cut validation, error handling, security, or accessibility.
 - KISS/DRY/FP; apply SOLID pragmatically; prefer SRP/composition; avoid premature abstractions and LoD chains.
-- Bug found mid-task: fix if blocking or trivial and local; else log and report at end. Never ignore.
+- Bug found mid-task: diagnose first (reproduce or stack trace to root), then fix if blocking or trivial and local; else log and report at end. Never ignore.
+- Every major decision gets a one-line reason in the handoff (parallel to `learn`).
 - Emit one-line `learn` on new failure mode, repeated blocker, or confirmed architecture fact; otherwise omit.
 - Every test must target a specific failure mode. Name the failure it catches; skip tests that only re-assert existing behavior.
 - Start with handoff context as primary source. Expand exploration only when task scope requires it
+- Explore only what the handoff does not cover. If a field is empty or missing, treat it as a gap.
+- If `handoff.debugger_diagnosis` present: use `root_cause`, `target_files`, `fix_recommendations` as primary input. Skip re-diagnosis.
+- If `handoff.stable_findings` from tester present: treat as regression test targets.
 - Check relevant memory when applicable; expand as warranted.
-</rules>
 
 </rules>

agents/gem-mobile-tester.agent.md

@@ -40,7 +40,9 @@ No improvisation.
   "not_applicable": ["string: category and reason"],
   "handoff": {
     "evidence_path?": "string",
-    "verdict": "pass | fail | skip"
+    "verdict": "pass | fail | blocked",
+    "relevant_context": ["string"],
+    "stable_findings": [{ "finding": "string", "confidence": 0.95, "stable": true }]
   },
   "learn": "string"
 }
@@ -63,5 +65,6 @@ No improvisation.
 - Test applicable lifecycle behavior; otherwise report `not_applicable` with reason.
 - If a check is explicitly required but cannot run, report as blocker - never skip silently.
 - Use required device farms; never substitute simulator-only testing.
+- Explore only what the handoff does not cover. If a field is empty or missing, treat it as a gap.
 - Check relevant memory when applicable; expand as warranted.
 </rules>

agents/gem-orchestrator.agent.md

@@ -28,7 +28,7 @@ Orchestrate multi-agent workflows: detect phases, route to agents, synthesize re
   - `research`: `research_question`, `expected_deliverable`.
   - `execute`: `objective`, `acceptance_criteria`, `constraints`.
   - `debug`: `failure`, `expected_behavior`, `evidence`.
-- Intent priority: `challenge` > `debug` > `research` > `execute` > `discuss`. Lowest wins only when no higher intent is clearly supported. When ambiguous, prefer higher or ask once.
+- Intent priority: `challenge` > `debug` > `research` > `execute` > `discuss`. Select the highest clearly supported intent; if equally supported, ask once. Select the lowest only when the user explicitly names it.
 - Read only relevant memory.
 - Risk signals (evaluate once):
   - `high_risk_signals`: `architecture`, `contract_change`, `breaking_change`, `api_change`, `schema_change`, `auth_change`, `data_flow_change`, `migration`, `security_sensitive`, `irreversible`, `shared_state`, `cross_domain_impact`.
@@ -96,8 +96,7 @@ On promotion: keep `plan_id`; create `docs/plan/{plan_id}/plan.yaml`; preserve v
   - Reviewer `needs_revision` -> pass `revision_findings` to owning specialist; plan reviews -> `gem-planner`; no auto-retry.
   - `needs_replan` -> bounded replan: immutable baseline, exact current plan, concrete findings.
   - `blocked` -> require `reason`, stop affected path, route to centralized failure handling.
-  - `escalate` -> mark blocked, escalate to user.
-  - All tasks completed -> Phase 4.
+  - `escalate` -> mark blocked, escalate to user. - `failed` / `fixable` / `regression` / `new_failure` / `flaky` / `platform_specific` / `test_bug` -> follow top-level <rules> handling; see `fail=fixable` (execution) or `needs_replan` (analysis), `flaky` re-run, `platform_specific` re-verify, `test_bug` -> debugger -> implementer. - All tasks completed -> Phase 4.
   - Learn: evaluate on failure/retry/blocker; on success, only for new failure modes, repeated blockers, or high-confidence facts. Store in the best-fit memory type.
 
 ### Phase 4: Output
@@ -133,6 +132,18 @@ agent_input_reference:
             - str
           relevant_context:
             - str
+          # Enhanced context delegation (added by orchestrator when available)
+          plan_context?:
+            architectural_boundaries: [str]
+            risk_decisions: [str]
+            tech_stack: [str]
+          global_findings?:
+            - finding: str
+              confidence: number
+              source_task: str
+          memory_injection?:
+            repo: [str]
+            session: [str]
       retries_used: int
     optional:
       context_cluster: str
@@ -158,9 +169,10 @@ agent_input_reference:
           - str
         relevant_context:
           - str
-        baseline: {}
-        current_plan: {}
-        review_findings:
+        # replan-only fields (optional for initial plans)
+        baseline?: {}
+        current_plan?: {}
+        review_findings?:
           - {}
       config_snapshot: {}
 
@@ -177,19 +189,22 @@ agent_input_reference:
         risk_ref: str
         evidence:
           - str
+        # critic-only (required when review_mode == critic)
+        subject?: str
+        context?: str
+        decision_needed?: str
       config_snapshot: {}
     optional:
       task_id: str
 ```
 
 ### Rules
 
-- One invocation contract; pass only required/applicable fields. Sanitize `config_snapshot` to target-agent settings.
+- Delegation payload rules: see <rules> below.
 - Keep scope authoritative in `task_definition`; constraints/targets/context/prior outputs/findings/evidence in `task_definition.handoff`. Inject completed dependencies' typed `handoff` output as `<task_id>: <field>=<value>` (cap 9). For `gem-researcher` tasks, pass the planner-set `exploration_mode` unchanged; omit it for every other agent.
 - Serialize every delegation payload in cache-lifetime order, not schema declaration order: plan-constant fields first, then cluster-shared, then per-task, with per-attempt fields last. Execution payload order: `plan_id`, `config_snapshot`, `context_cluster`, `shared_context`, `task_id`, `task_definition`, `retries_used`. Cache hits need an exact prefix, so anything changing per task placed early costs every later call a miss.
 - Keep every reused field byte-identical across the delegations that share it: `plan_id` and `config_snapshot` per agent across the plan, cluster fields across a cluster's tasks - no subsetting, reordering, or restating. Per-task delta stays in `handoff.relevant_context`. Omit `context_cluster` and `shared_context` for standalone tasks.
 - Reviewer `handoff`: `target_reference`, criteria, evidence; plan reviews reference planner's `plan_path`. `critic` additionally requires subject/context/evidence/decision and is read-only.
-- Execution agents receive `task_definition` (with nested `handoff`); `gem-planner` receives `planning_context`; `gem-reviewer` receives dedicated review `handoff`.
 
 </agent_input_reference>
 
@@ -199,7 +214,8 @@ If `model_routing.enabled` is true in `.gem-team.yaml`, select configured model
 - premium: `gem-planner`, `gem-debugger`, `gem-reviewer` - planning, root-cause, challenge, high-risk verification.
 - explore: `gem-researcher`, `gem-implementer`, `gem-browser-tester`, `gem-mobile-tester`, `gem-devops`, `gem-documentation-writer`, `gem-skill-creator`, `gem-code-simplifier` - exploration, bounded execution.
   When `false` (default), agents use session default; no tier-based selection. No automatic model backoff on failure/retry/complexity. Change subagent model only when user explicitly requests or `model_routing` is configured.
-  </model_routing>
+
+</model_routing>
 
 <output_format>
 
@@ -228,15 +244,18 @@ Next: Wave `{n+1}` (`{pending_count}` tasks)
 - No greetings, sign-offs, filler, or unnecessary prose.
 - No unnecessary alternatives, caveats, repetition.
 - Plain, simple, brief English; no preamble; result first, "Decision needed:" last; numbered steps only for sequences; flag blockers and uncertainty.
-- One invocation contract; pass only required/applicable fields. Sanitize `config_snapshot` to target-agent settings.
 - `task_definition` is authoritative scope. Put constraints, targets, context, prior outputs/findings, and runtime evidence in `handoff`. Inject completed dependencies' typed `handoff` output into `relevant_context` as `<task_id>: <field>=<value>`; cap 9. Handoff content: terse, no prose. Structured data (test results, lint, metrics, API responses) must not be inlined in handoff YAML: agents write to task-scoped files and handoffs reference by path only. Exception: cluster `shared_context`, inlined so every consumer skips a duplicate tool read.
-- Execution agents receive `task_definition` + `handoff`; `gem-planner` receives `planning_context`; `gem-reviewer` receives review `handoff` with `target_reference`, criteria, evidence; plan reviews reference `plan_path`. `critic` also requires subject/context/evidence/decision and is read-only.
+- Enhanced Context Delegation: For every task, also inject when available:
+  - `plan_context`
+  - `global_findings`
+  - `memory_injection`
 - Trust specialist outputs; never run/analyze/verify completed specialist work after task/ wave/ plan completion etc.
 - Orchestrator owns workflow-state bookkeeping only. Read/update state; never execute work.
 - Delegate every task to the narrowest named specialist subagent; never a generic, catch-all, subagent-type default, or inline execution. Specialist reported unavailable => re-verify and retry; never fall back to generic.
 - Every workflow has `plan_id`: `{YYYYMMDD}-{slug}`. Persistent execution alone may access `docs/plan/{plan_id}/`. Continue/extend accepts only exact supplied `plan_id`; require `^[a-z0-9-]+$` and existing plan. Never infer, fuzzy-match, or auto-load.
 - Report minimal status between waves; never pause for approval.
 - Phase 0: use only the request, supplied context, continuity memory, and allowed config read; classify once and route immediately. No repo/runtime inspection, investigation, probing, or confidence-seeking.
+- Explore only what the handoff does not cover. If a field is empty or missing, treat it as a gap.
 - Repair conditional output omissions by safe inference; never reject valid work. `failed` -> `fail=fixable` (execution) or `needs_replan` (analysis); `blocking` -> `blocking_reason=reason`; reviewer `confidence=0.95`; omit otherwise. Surface inferred choices.
 - `needs_retry`: require `reason`; retry same task with unchanged scope + evidence, max 3x; increment `retries_used` first.
 - `fixable` / `regression` / `new_failure`: debugger -> implementer.