feat(shared,extension): add next-step review feedback

Persist local review feedback, suppress dismissed or snoozed review items across dashboard and Chickens, and record the corresponding feature-pack/design/config updates.

Author: Oba-One

.codex/config.toml

@@ -2,7 +2,7 @@
 # See also: AGENTS.md (shared instructions), .claude/ (Claude Code config)
 
 # Model defaults
-model = "gpt-5.4"
+model = "gpt-5.5"
 
 # Sandbox: keep work inside the repo by default; approval covers anything broader.
 sandbox_mode = "workspace-write"
@@ -25,7 +25,7 @@ multi_agent = true
 model = "gpt-5.4"
 model_reasoning_effort = "high"
 sandbox_mode = "read-only"
-approval_policy = "never"
+approval_policy = "on-request"
 
 [profiles.quick]
 model = "gpt-4.1-mini"
@@ -35,13 +35,13 @@ model_reasoning_effort = "low"
 model = "gpt-5.4"
 model_reasoning_effort = "high"
 sandbox_mode = "read-only"
-approval_policy = "never"
+approval_policy = "on-request"
 
 [profiles.triage]
 model = "gpt-4.1-mini"
 model_reasoning_effort = "low"
 sandbox_mode = "read-only"
-approval_policy = "never"
+approval_policy = "on-request"
 
 [profiles.migrate]
 model = "gpt-5.4"

.plans/features/agent-knowledge-sandbox/eval/qa-report.md

@@ -12,6 +12,8 @@
 - Retrieval and provenance surfaces need an honest regression pass before this pack can move beyond
   blocked QA.
 - Earlier plan notes described a more ambitious graph backend than the one currently implemented.
+- QA must also verify the newer memory-provenance contract: observed, inferred, user-confirmed,
+  imported, and stale memories should not collapse into the same retrieval or UI behavior.
 
 ## QA Pass 1: Codex
 

.plans/features/agent-knowledge-sandbox/qa/qa-codex.todo.md

@@ -17,7 +17,7 @@ skills:
 qa_order: 1
 handoff_in: handoff/qa-codex/agent-knowledge-sandbox
 handoff_out: handoff/qa-claude/agent-knowledge-sandbox
-updated: 2026-04-19
+updated: 2026-05-07
 ---
 
 # QA Pass 1 — Codex
@@ -75,10 +75,14 @@ Triggered by: UI + State lanes complete
 - [ ] Positive precedents boost confidence >= 0.05
 - [ ] Negative precedents decrease confidence >= 0.05
 - [ ] Trace quota enforcement (max 500) works with oldest-first pruning
+- [ ] Memory write-back records provenance label, confirmation status, source channel, provider/model use, trace/task ID, confidence, and unresolved questions where applicable
+- [ ] Inferred but unconfirmed memory is retrievable as context but is not treated as instruction-like guidance
+- [ ] Stale memory is visibly labeled and ranked below current observed or user-confirmed context
 
 ## Integration
 
 - [ ] All existing unit tests pass (zero regressions)
 - [ ] All existing eval cases pass at current thresholds
 - [ ] Agent cycle time within 120% of baseline
 - [ ] Flat agentMemories table still works (backwards compat)
+- [ ] Recommendation UI can answer "where did this come from?" without exposing raw prompts, model output, or internal trace payloads in simple mode

.plans/features/agent-knowledge-sandbox/spec.md

@@ -19,6 +19,9 @@ The "YouTube Kids for agents" model: sandbox the knowledge, not just the executi
 - The agent has no way to learn from YouTube channels, GitHub repos, or RSS feeds that the coop follows
 - Reasoning traces evaporate after skill execution — no institutional memory, no precedent system
 - The Neo4j Context Graph talk validated that graph memory + reasoning traces + hybrid retrieval is production-ready and the patterns map directly to Coop's architecture
+- The durable-agent runtime lesson is clear: if models can change, memory needs independent
+  provenance, confirmation status, and retrieval rules so Coop behaves continuously without
+  turning model output into unaccountable truth.
 
 ## Scope
 
@@ -57,7 +60,7 @@ The "YouTube Kids for agents" model: sandbox the knowledge, not just the executi
 - Context assembly for skill prompts (token-budgeted)
 - No LLM calls during retrieval (hard requirement)
 
-**Phase 6 — Reasoning Traces + Compound Loop** (learning)
+**Phase 6 — Reasoning Traces + Memory Provenance** (learning)
 - Record decision traces as precedent nodes linked to skill runs
 - Precedent query by observation similarity
 - Confidence adjustment based on past decision outcomes
@@ -66,6 +69,14 @@ The "YouTube Kids for agents" model: sandbox the knowledge, not just the executi
 - Rejection weakens: decrease edge confidence, temporal invalidation (not deletion)
 - Validated insight entities: approved draft summaries become first-class graph nodes
 - Append-only activity log: chronological record of ingests, queries, lint passes
+- Memory write-back labels each durable entry as `observed`, `inferred`, `user-confirmed`,
+  `imported`, or `stale`
+- Confirmation status controls retrieval weight and whether a memory can be used as instruction-like
+  context later
+- Retrieval-before-work gathers relevant sources, decisions, prior failures, open questions, and
+  constraints before meaningful agent work starts
+- Write-back-after-work records output summary, source channel, provider/model use, trace or task ID,
+  confidence, unresolved questions, and confirmation status
 
 **Phase 7 — Lint + Integration** (wiring + health)
 - Knowledge lint skill: orphan entities, stale sources, contradictions, coverage gaps, graph health
@@ -91,12 +102,14 @@ The "YouTube Kids for agents" model: sandbox the knowledge, not just the executi
 - Add YouTube channels, GitHub repos, RSS feeds, subreddits, NPM packages as knowledge sources
 - See what the agent knows (topic bars) and why it recommended something (sourced from + track record)
 - Review agent decisions with full provenance (which sources, which precedents)
+- Distinguish user-confirmed memory from inferred or imported memory without opening an agent log
 - See source health at a glance (popup dot, Nest freshness indicators)
 
 **Operators can:**
 - Configure source allowlists per coop
 - Monitor graph size and entity counts
 - Review agent decision history with reasoning traces
+- Audit which memories are model-inferred, source-observed, imported, stale, or confirmed by a member
 - See cascade effects before removing sources
 
 **What stays the same:**
@@ -118,6 +131,10 @@ The "YouTube Kids for agents" model: sandbox the knowledge, not just the executi
 - Entity extraction must use existing inference cascade — no new model infrastructure
 - No LLM calls during graph retrieval (performance requirement)
 - Source adapters must go through `assertAllowedSource()` — no direct fetch from unapproved URLs
+- Provenance labels extend the existing memory/trace contracts; do not introduce a parallel
+  repo-level memory truth surface.
+- Raw source fetch state stays local. Only approved outputs and user-confirmed memory projections may
+  become shared coop memory.
 
 ### New dependencies
 - `@kuzu/kuzu-wasm` — embedded graph DB (or Vela-Engineering fork for concurrent writes)
@@ -192,18 +209,25 @@ The "YouTube Kids for agents" model: sandbox the knowledge, not just the executi
 - [ ] Precedent query finds similar past decisions
 - [ ] Positive precedents boost confidence >= 0.05
 - [ ] Negative precedents decrease confidence >= 0.05
+- [ ] Memory write-back records provenance label, confirmation status, source channel, provider or
+  model use, trace/task ID, confidence, and unresolved questions where applicable
+- [ ] Unconfirmed inferred memory is retrievable as context but never elevated to instruction-like
+  guidance without member confirmation
+- [ ] Stale memory is visible as stale and ranked below current observed or confirmed context
 
 ### Phase 7 — Integration
 - [ ] All existing eval cases pass at or above current thresholds (zero regression)
 - [ ] Agent cycle time stays within 120% of baseline
 - [ ] Graph-enhanced skills show >= 10% quality improvement (A/B evaluation)
 - [ ] UI surfaces work: Nest Sources, Roost Knowledge, DraftCard provenance, Popup pulse
+- [ ] Recommendation surfaces can answer "where did this come from?" without exposing raw prompts,
+  model output, or internal trace payloads in simple mode
 
 ## Validation Plan
 
 - **Unit**: Source registry CRUD, adapter parsing, entity extraction, graph CRUD, retrieval relevance, temporal correctness, reasoning traces
-- **Integration**: Full pipeline: source → adapter → extraction → graph → retrieval → skill context → output
-- **E2E**: Member adds source → agent ingests → agent uses in recommendation → member sees provenance
+- **Integration**: Full pipeline: source → adapter → extraction → graph → retrieval → skill context → output → provenance-labeled write-back
+- **E2E**: Member adds source → agent ingests → agent uses in recommendation → member sees provenance and can distinguish inferred from confirmed memory
 - **A/B**: Baseline (flat memory) vs graph-enhanced (graph retrieval) quality comparison on eval corpus
 - **Regression**: All existing skill eval cases + unit tests must pass at pre-implementation thresholds
 

.plans/features/agent-knowledge-sandbox/status.json

@@ -60,8 +60,9 @@
     "No LLM during retrieval (hard perf requirement)",
     "Vellum material language for all UI surfaces",
     "7-phase dependency-ordered build with gates",
-    "Entity extraction as new skill using existing cascade"
+    "Entity extraction as new skill using existing cascade",
+    "Memory provenance and confirmation labels are part of the product contract: observed, inferred, user-confirmed, imported, and stale memories must behave differently in retrieval and UI."
   ],
-  "updated_at": "2026-04-19",
-  "notes": "UI + state lanes materially landed; graph backend still snapshot-persisted (Kuzu-WASM deferred). QA pass 1 is now ready to run."
+  "updated_at": "2026-05-07",
+  "notes": "UI + state lanes materially landed; graph backend still snapshot-persisted (Kuzu-WASM deferred). QA pass 1 is now ready to run, including provenance, confirmation, and retrieval-before-work checks."
 }

.plans/features/next-gen-model-readiness/context.md

@@ -8,6 +8,24 @@ Analysis based on Nate B Jones transcript re: Claude Mythos and four parallel au
 3. Validation pipeline redundancy analysis
 4. Multi-agent coordination overhead assessment
 
+Additional source: "Dive into Claude Code: The Design Space of Today's and Future AI Agent
+Systems" (arXiv:2604.14228v1). The relevant takeaway is not to copy Claude Code wholesale; it is
+to preserve deterministic harness boundaries while simplifying model-facing scaffolding.
+
+## Harness Guardrail Classification
+
+Use this table during Phase 1 cleanup before removing or rewriting prompt/context material.
+
+| Classification | Examples in this repo | Where it must live |
+|---|---|---|
+| `deterministic-gate` | permission checks, deny/allow rules, hook enforcement, schema validation, `bun run test` enforcement, release gates | Code, hook config, schemas, validators, tests, or `scripts/validate.ts` |
+| `repo-constraint` | barrel imports, root `.env.local`, no Dexie access from views, MV3 service-worker constraints | One canonical repo instruction or rule file |
+| `product-intent` | local-first, passkey-first, explicit publish, community/project framing, friendly non-console UX | `CLAUDE.md`, product context, or current plan spec |
+| `soft-guidance` | library tutorials, static file maps, generic debugging recipes, boilerplate test snippets | Remove or replace with file pointers |
+
+If a cleanup step touches a `deterministic-gate`, the implementation note must name the executable
+surface that still enforces it after the prose is removed.
+
 ## Key Files by Phase
 
 ### Phase 1: Prompt Surface

.plans/features/next-gen-model-readiness/lanes/api.claude.todo.md

@@ -21,14 +21,17 @@ done_when:
 skills:
   - architecture
   - testing
-updated: 2026-04-02
+updated: 2026-05-07
 ---
 
 # Phase 3: Prepare the Agent Pipeline for Model Upgrade
 
 Target: Introduce a "capable model" code path alongside the 0.5B legacy path. Define tools from deterministic skills. Collapse output handlers into generic tools. Feature-flagged via `VITE_COOP_AGENT_MODE`.
 
-**Principle**: The legacy path (0.5B + heuristic fallbacks) continues working unchanged. The autonomous path is additive — same observation lifecycle, same approval gates, same memory system, different execution strategy.
+**Principle**: The legacy path (0.5B + heuristic fallbacks) continues working unchanged. The
+autonomous path is additive — same observation lifecycle, same approval gates, same memory system,
+same trace evidence, same fallback semantics, different execution strategy. Model routing remains
+internal runtime evidence; simple mode should not become a provider-management surface.
 
 ## Step 1: Add `VITE_COOP_AGENT_MODE` environment variable
 
@@ -162,6 +165,8 @@ export async function runAutonomousAgentCycle(options: {
 - The plan goes through the same approval gate as the legacy path
 - Memory is queried the same way (via `queryMemoriesForSkill`)
 - Observations are created/updated the same way
+- Provider/model details are persisted for traces, benchmarks, and advanced diagnostics, but user
+  workflow state stays provider-independent
 
 **Model bridge interface** (abstract — concrete implementation depends on model provider):
 ```typescript

.plans/features/next-gen-model-readiness/lanes/docs.claude.todo.md

@@ -22,13 +22,25 @@ done_when:
   - rules/tests.md lines < 100
 skills:
   - architecture
-updated: 2026-04-02
+updated: 2026-05-07
 ---
 
 # Phase 1: Simplify the Prompt Surface
 
 Target: ~8,500 lines → ~3,000 lines. Remove library docs, static code maps, procedural recipes, duplicated constraints. Keep outcomes, constraints, anti-patterns, product intent.
 
+## Required guardrail
+
+Before deleting or shortening any instruction, classify it using `../context.md`:
+
+- `deterministic-gate`: keep or move to code, hooks, schemas, validators, tests, or validation scripts.
+- `repo-constraint`: keep once in the canonical repo instruction/rule surface.
+- `product-intent`: preserve in product context or another short pointer.
+- `soft-guidance`: remove or replace with a source pointer.
+
+Do not convert a deterministic gate into prompt-only guidance. If the old prose was the only place a
+gate existed, stop and route it to an executable surface or record the gap.
+
 ## Step 1: Replace context code maps with pointer files
 
 Replace `.claude/context/app.md` (163 lines), `extension.md` (466 lines), `shared.md` (448 lines) with ~15-line pointer files. Each file becomes:
@@ -53,6 +65,9 @@ Read the source files above for architecture details.
 
 **Verify**: Each pointer file < 20 lines. No constraint lost (cross-reference against rules/ files).
 
+**Guardrail audit**: Record any removed `deterministic-gate`, `repo-constraint`, or `product-intent`
+in the implementation notes with its new canonical home.
+
 ## Step 2: Reduce skills to constraint cards
 
 For each `.claude/skills/*/SKILL.md`, reduce to 30-50 lines:
@@ -93,6 +108,9 @@ For each `.claude/skills/*/SKILL.md`, reduce to 30-50 lines:
 
 **Verify**: `wc -l .claude/skills/*/SKILL.md` shows each file 20-50 lines.
 
+**Guardrail audit**: For each skill, keep Coop-specific constraints and anti-patterns; remove generic
+methodology only after confirming enforcement-sensitive behavior lives in a rule, hook, schema, or test.
+
 ## Step 3: Delete meta-documentation
 
 - [ ] Delete `.claude/skills/index.md` (238 lines) — meta-documentation about the prompt system itself. The model discovers skills via registry, not an index file.
@@ -139,6 +157,9 @@ Remove ~120 lines from CLAUDE.md:
 
 **Verify**: `wc -l CLAUDE.md` shows < 200 lines. `bun run validate quick` passes.
 
+**Guardrail audit**: Any removed release, permission, env, or command rule must still be enforced by a
+hook, rule file, validator, test, or script.
+
 ## Step 6: Deduplicate constraints
 
 For each constraint in the deduplication map (see context.md), ensure it appears in exactly one canonical location:
@@ -166,4 +187,5 @@ Review each `.claude/agents/*.md` for duplicated Coop rules:
 - [ ] `bun run validate quick` passes
 - [ ] Total `.claude/` line count < 3,500 (down from ~8,500)
 - [ ] No constraint was removed without being present elsewhere (audit trail in commit message)
+- [ ] Every removed `deterministic-gate`, `repo-constraint`, and `product-intent` has an implementation-note entry naming its durable home
 - [ ] `product.md` unchanged