diff --git a/.github/actions/agent-context-store/action.yml b/.github/actions/agent-context-store/action.yml
deleted file mode 100644
index b90b197ea..000000000
--- a/.github/actions/agent-context-store/action.yml
+++ /dev/null
@@ -1,208 +0,0 @@
-name: 'Agent Context Store'
-description: >
-  Load / save durable *work context* for agentic GitHub Actions runs.
-  Never stores tokens, cookies, PoW answers, or raw conversation bodies.
-  Auth sessions stay in $RUNNER_TEMP and are discarded (see #112/#114 policy).
-
-inputs:
-  mode:
-    description: 'load | save'
-    required: true
-  context_key:
-    description: >
-      Stable key for this work unit (e.g. pr-123-review, issue-118-triage).
-      Used as Actions cache key suffix.
-    required: true
-  context_json:
-    description: 'JSON string to save (mode=save). Must be non-secret fields only.'
-    required: false
-    default: '{}'
-  github_token:
-    description: 'Token for optional comment-marker fallback (GITHUB_TOKEN is fine)'
-    required: false
-    default: ''
-  target_number:
-    description: 'PR or issue number for comment-marker bookmark (optional)'
-    required: false
-    default: ''
-  repository:
-    description: 'owner/repo'
-    required: false
-    default: ''
-
-outputs:
-  context_json:
-    description: 'Loaded context JSON (mode=load); empty object if miss or mode=save'
-    value: ${{ steps.load.outputs.context_json || '{}' }}
-  cache_hit:
-    description: 'true if a previous context file was restored (exact or via restore-keys)'
-    value: ${{ steps.hit.outputs.cache_hit || 'false' }}
-
-runs:
-  using: composite
-  steps:
-    - name: Resolve cache paths
-      id: paths
-      shell: bash
-      env:
-        CONTEXT_KEY: ${{ inputs.context_key }}
-        MODE: ${{ inputs.mode }}
-      run: |
-        set -euo pipefail
-        if [[ "$MODE" != "load" && "$MODE" != "save" ]]; then
-          echo "::error::mode must be 'load' or 'save' (got: $MODE)"
-          exit 1
-        fi
-        CTX_DIR="${RUNNER_TEMP}/agent-context"
-        mkdir -p "$CTX_DIR"
-        KEY_SAFE=$(printf '%s' "$CONTEXT_KEY" | tr -c 'A-Za-z0-9._-' '_')
-        KEY_HASH=$(printf '%s' "$CONTEXT_KEY" | sha256sum | cut -c1-8)
-        KEY_SAFE="${KEY_SAFE}-${KEY_HASH}"
-        echo "dir=${CTX_DIR}" >> "$GITHUB_OUTPUT"
-        echo "file=${CTX_DIR}/${KEY_SAFE}.json" >> "$GITHUB_OUTPUT"
-        echo "key_safe=${KEY_SAFE}" >> "$GITHUB_OUTPUT"
-
-    - name: Restore context cache
-      id: cache
-      uses: actions/cache@v4
-      with:
-        path: ${{ steps.paths.outputs.dir }}
-        key: agent-ctx-${{ runner.os }}-${{ steps.paths.outputs.key_safe }}-${{ github.run_id }}-${{ github.run_attempt }}
-        restore-keys: |
-          agent-ctx-${{ runner.os }}-${{ steps.paths.outputs.key_safe }}-
-
-    - name: Detect restored context
-      id: hit
-      shell: bash
-      env:
-        FILE: ${{ steps.paths.outputs.file }}
-      run: |
-        set -euo pipefail
-        if [ -f "$FILE" ]; then
-          echo "cache_hit=true" >> "$GITHUB_OUTPUT"
-        else
-          echo "cache_hit=false" >> "$GITHUB_OUTPUT"
-        fi
-
-    - name: Load context
-      id: load
-      if: inputs.mode == 'load'
-      shell: bash
-      env:
-        FILE: ${{ steps.paths.outputs.file }}
-        CONTEXT_KEY: ${{ inputs.context_key }}
-      run: |
-        set -euo pipefail
-        if [ -f "$FILE" ]; then
-          python3 - <<'PY'
-        import json, os
-
-        def strip_forbidden(obj):
-            FORBIDDEN = {"token", "cookie", "cookies", "authorization", "api_key",
-                         "pow_answer", "pow_signature", "password", "secret"}
-            if isinstance(obj, dict):
-                return {
-                    k: strip_forbidden(v)
-                    for k, v in obj.items()
-                    if k.lower() not in FORBIDDEN
-                    and not any(x in k.lower() for x in ("token", "cookie", "secret", "password", "pow_"))
-                }
-            if isinstance(obj, list):
-                return [strip_forbidden(i) for i in obj]
-            return obj
-
-        path = os.environ["FILE"]
-        try:
-            with open(path) as f:
-                data = json.load(f)
-        except (json.JSONDecodeError, OSError) as e:
-            print(f"::warning::corrupt or unreadable context ({e}); discarding")
-            data = {}
-        if not isinstance(data, dict):
-            print("::warning::context is not an object; discarding")
-            data = {}
-        cleaned = strip_forbidden(data)
-        raw = json.dumps(cleaned, separators=(",", ":"))
-        if len(raw) > 32_000:
-            print("::warning::context truncated to 32k")
-            cleaned = {"_truncated": True, "keys": list(cleaned.keys())[:20]}
-            raw = json.dumps(cleaned, separators=(",", ":"))
-        with open(os.environ["GITHUB_OUTPUT"], "a") as out:
-            out.write(f"context_json={raw}\n")
-        print(f"::notice::Loaded context keys={list(cleaned.keys())}")
-        PY
-        else
-          echo 'context_json={}' >> "$GITHUB_OUTPUT"
-          echo "::notice::No prior context for key=${CONTEXT_KEY}"
-        fi
-
-    - name: Save context
-      if: inputs.mode == 'save'
-      shell: bash
-      env:
-        INPUT_JSON: ${{ inputs.context_json }}
-        FILE: ${{ steps.paths.outputs.file }}
-        CONTEXT_KEY: ${{ inputs.context_key }}
-      run: |
-        set -euo pipefail
-        python3 - <<'PY'
-        import json, os, time
-
-        def strip_forbidden(obj):
-            FORBIDDEN = {"token", "cookie", "cookies", "authorization", "api_key",
-                         "pow_answer", "pow_signature", "password", "secret"}
-            if isinstance(obj, dict):
-                return {
-                    k: strip_forbidden(v)
-                    for k, v in obj.items()
-                    if k.lower() not in FORBIDDEN
-                    and not any(x in k.lower() for x in ("token", "cookie", "secret", "password", "pow_"))
-                }
-            if isinstance(obj, list):
-                return [strip_forbidden(i) for i in obj]
-            return obj
-
-        raw_in = os.environ.get("INPUT_JSON") or "{}"
-        try:
-            data = json.loads(raw_in)
-        except json.JSONDecodeError:
-            data = {}
-        if not isinstance(data, dict):
-            data = {}
-        cleaned = strip_forbidden(data)
-        cleaned["_saved_at"] = time.strftime("%Y-%m-%dT%H:%M:%SZ", time.gmtime())
-        cleaned["_key"] = os.environ.get("CONTEXT_KEY", "")
-        raw = json.dumps(cleaned, separators=(",", ":"))
-        if len(raw) > 32_000:
-            print("::warning::context exceeds 32k on save; truncating")
-            cleaned = {"_truncated": True, "_saved_at": cleaned["_saved_at"], "_key": cleaned["_key"], "keys": list(cleaned.keys())[:20]}
-        path = os.environ["FILE"]
-        os.makedirs(os.path.dirname(path), exist_ok=True)
-        with open(path, "w") as f:
-            json.dump(cleaned, f, indent=2)
-        print(f"::notice::Saved agent context → {path} keys={list(cleaned.keys())}")
-        PY
-
-    - name: Optional comment-marker bookmark
-      if: inputs.mode == 'save' && inputs.target_number != '' && inputs.github_token != ''
-      shell: bash
-      env:
-        GH_TOKEN: ${{ inputs.github_token }}
-        REPO: ${{ inputs.repository }}
-        TARGET: ${{ inputs.target_number }}
-        KEY: ${{ inputs.context_key }}
-      run: |
-        set -euo pipefail
-        if ! [[ "${TARGET:-}" =~ ^[0-9]+$ ]]; then exit 0; fi
-        if [ -z "${REPO:-}" ]; then REPO="${{ github.repository }}"; fi
-        MARKER="<!-- agent-ctx:${KEY} -->"
-        # gh api has no --arg; pipe to jq for safe filtering
-        EXISTING=$(gh api --paginate "repos/${REPO}/issues/${TARGET}/comments" 2>/dev/null \
-          | jq -r --arg m "$MARKER" '.[] | select(.body | contains($m)) | .id' \
-          | head -1 || true)
-        BODY=$(printf '%s\n_Context bookmark for `%s` (non-secret). Auth sessions remain ephemeral._\n' "$MARKER" "$KEY")
-        if [ -n "$EXISTING" ]; then
-          gh api -X PATCH "repos/${REPO}/issues/comments/${EXISTING}" -f body="$BODY" >/dev/null || true
-        else
-          gh api -X POST "repos/${REPO}/issues/${TARGET}/comments" -f body="$BODY" >/dev/null || true
-        fi
diff --git a/.github/connectors/README.md b/.github/connectors/README.md
deleted file mode 100644
index a594bc90f..000000000
--- a/.github/connectors/README.md
+++ /dev/null
@@ -1,39 +0,0 @@
-# Connector Catalog (AUTHORIZED)
-
-Centralized, non-secret configuration for external integrations.
-Runtime code and Actions read **environment variable names** only — never secret values.
-
-## Authorized secrets (GitHub Actions / repo secrets)
-
-| Secret name | Purpose | Used by |
-|-------------|---------|--------|
-| `OPERATOR_GITHUB_TOKEN` | Write/admin PAT for CodeRabbit autofix, privileged comments, continuous-ops | peer-review-orchestrator, continuous-ops, auto-jules |
-| `GITHUB_TOKEN` | Default Actions token (contents/PRs/issues scoped) | all workflows |
-| `OMNI_API_KEY` / `OMNIROUTE_API_KEY` | OmniRoute free-tier peer | model-router, http-llm-invoke |
-| `OMNI_BASE_URL` | Optional override (must be https allow-listed host) | http-llm-invoke |
-| `OPENROUTER_API_KEY` | OpenRouter free-tier peer (`:free` models only) | model-router, http-llm-invoke |
-| `GEMINI_API_KEY` | Gemini residual path | gemini-* workflows |
-| Linear / Vercel / other | Connected via Grok Connectors or separate secrets as needed | agent-feedback-linear-sync, deploy |
-
-## Provider allow-list (http-llm-invoke)
-
-- `cloud.omniroute.online`
-- `openrouter.ai` / `api.openrouter.ai`
-
-HTTPS only. Non-allow-listed hosts are rejected.
-
-## Soft-budget policy (exceed prior capacity)
-
-See `scripts/model_router.py` and `docs/schemas/model-success-matrix.yaml`.
-Omni `auto/best-free` and OpenRouter free models have elevated daily soft limits
-so peer capacity is preferred over Gemini residual. Limits are best-effort
-(Actions cache per Pacific day).
-
-## Files in this directory
-
-- `github.yaml` — repo identity + API surface
-- `llm-peers.yaml` — Omni / OpenRouter / Gemini routing preferences
-- `integrations.yaml` — Linear, Vercel, Jules, CodeRabbit, Devin markers
-
-Exchange connectors (Yobit/KuCoin/Binance) were **intentionally removed** from
-the active surface (PR #74 closed). Do not reintroduce without OPERATOR sign-off.
diff --git a/.github/connectors/github.yaml b/.github/connectors/github.yaml
deleted file mode 100644
index de09cf291..000000000
--- a/.github/connectors/github.yaml
+++ /dev/null
@@ -1,14 +0,0 @@
-github:
-  repository:
-    owner: timerloggedout-spec
-    name: termux-monorepo
-  api:
-    base_url: https://api.github.com
-    token_env: GITHUB_TOKEN
-    operator_token_env: OPERATOR_GITHUB_TOKEN
-  permissions_expected:
-    - contents: read|write
-    - pull-requests: write
-    - issues: write
-    - actions: read
-    - checks: read
diff --git a/.github/connectors/integrations.yaml b/.github/connectors/integrations.yaml
deleted file mode 100644
index 2b38f2b9c..000000000
--- a/.github/connectors/integrations.yaml
+++ /dev/null
@@ -1,19 +0,0 @@
-integrations:
-  coderabbit:
-    autofix_trigger: "@coderabbitai autofix"
-    requires_operator_token: true
-  jules:
-    primary_builder: true
-    async_cloud_vm: true
-  devin:
-    review_marker: true
-  linear:
-    workflow: agent-feedback-linear-sync.yml
-  vercel:
-    status_context: Vercel
-  peer_reviewers:
-    - coderabbit
-    - devin-ai
-    - aikido
-    - sentry
-    - copilot-pull-request
diff --git a/.github/connectors/llm-peers.yaml b/.github/connectors/llm-peers.yaml
deleted file mode 100644
index f9dbf52f7..000000000
--- a/.github/connectors/llm-peers.yaml
+++ /dev/null
@@ -1,57 +0,0 @@
-# Peer policy: Omni \u2194 OpenRouter nearly equal free-tier peers.
-# Selection = desired model + soft-budget headroom. Gemini = residual only.
-
-peers:
-  omni:
-    enabled_env: OMNI_API_KEY  # or OMNIROUTE_API_KEY
-    base_url_env: OMNI_BASE_URL
-    default_base: https://cloud.omniroute.online/v1
-    models:
-      - id: auto/best-free
-        roles: [triage, review, invoke]
-        soft_limits:
-          triage: 400
-          review: 250
-          invoke: 400
-  openrouter:
-    enabled_env: OPENROUTER_API_KEY
-    default_base: https://openrouter.ai/api/v1
-    require_free_suffix: true
-    models:
-      - id: google/gemma-4-31b-it:free
-        roles: [triage, review, invoke]
-        soft_limits: { triage: 80, review: 80, invoke: 80 }
-      - id: google/gemma-4-26b-a4b-it:free
-        roles: [triage, review, invoke]
-        soft_limits: { triage: 80, review: 80, invoke: 80 }
-      - id: meta-llama/llama-3.3-70b-instruct:free
-        roles: [triage, review, invoke]
-        soft_limits: { triage: 80, review: 80, invoke: 80 }
-      - id: google/gemma-3-12b-it:free
-        roles: [triage, review, invoke]
-        soft_limits: { triage: 80, review: 80, invoke: 80 }
-      - id: qwen/qwen3-coder:free
-        roles: [review, invoke]
-        soft_limits: { triage: 60, review: 60, invoke: 60 }
-      - id: deepseek/deepseek-r1:free
-        roles: [review]
-        soft_limits: { triage: 40, review: 40, invoke: 40 }
-      - id: cohere/north-mini-code:free
-        roles: [review, invoke]
-        soft_limits: { triage: 60, review: 60, invoke: 60 }
-  gemini:
-    enabled_env: GEMINI_API_KEY
-    residual_only: true
-    models:
-      - id: gemini-3.1-flash-lite
-        soft_limits: { triage: 450, review: 450, invoke: 450 }
-      - id: gemini-3.5-flash-lite
-        soft_limits: { triage: 450, review: 450, invoke: 450 }
-      - id: gemini-2.5-flash-lite
-        soft_limits: { triage: 20, review: 20, invoke: 20 }
-      - id: gemini-3.5-flash
-        soft_limits: { triage: 20, review: 20, invoke: 20 }
-      - id: gemini-2.5-flash
-        soft_limits: { triage: 20, review: 20, invoke: 20 }
-      - id: gemini-3-flash
-        soft_limits: { triage: 20, review: 20, invoke: 20 }
diff --git a/.github/workflows/agent-continuous-ops.yml b/.github/workflows/agent-continuous-ops.yml
deleted file mode 100644
index 5ece9aa90..000000000
--- a/.github/workflows/agent-continuous-ops.yml
+++ /dev/null
@@ -1,219 +0,0 @@
-name: Continuous agent ops
-
-# 24/7 unattended advancement of open agent PRs.
-# Complements event-driven agent-review-auto-jules + peer-review-orchestrator:
-# those fire on bot feedback; this catches PRs that went quiet, got dirty,
-# or never received a follow-up after peer-review-ready.
-#
-# Requires: Jules GitHub App (repo access). Optional: secrets.JULES_API_KEY
-# Security: no secrets in prompts; OPERATOR_TOKEN only for comment write if set.
-
-on:
-  schedule:
-    # Every 2 hours — enough to clear lag without burning review quota
-    - cron: '17 */2 * * *'
-  workflow_dispatch:
-    inputs:
-      force_all:
-        description: 'Ignore debounce and re-ping all eligible PRs'
-        required: false
-        default: 'false'
-      max_prs:
-        description: 'Max PRs to act on this run'
-        required: false
-        default: '8'
-
-concurrency:
-  group: continuous-agent-ops
-  cancel-in-progress: false
-
-permissions: {}
-
-jobs:
-  sweep:
-    runs-on: ubuntu-latest
-    timeout-minutes: 10
-    permissions:
-      contents: read
-      pull-requests: write
-      issues: write
-    steps:
-      - name: Sweep open PRs and advance stuck work
-        uses: actions/github-script@f28e40c7f34bde8b3046d885e986cb6290c5673b # v7.1.0
-        env:
-          FORCE_ALL: ${{ github.event.inputs.force_all || 'false' }}
-          MAX_PRS: ${{ github.event.inputs.max_prs || '8' }}
-          # Prefer operator token so @jules / @coderabbitai accept write-path comments
-          HAS_OPERATOR: ${{ secrets.OPERATOR_GITHUB_TOKEN != '' || secrets.OPERATOR_TOKEN != '' }}
-        with:
-          github-token: ${{ secrets.OPERATOR_GITHUB_TOKEN || secrets.OPERATOR_TOKEN || secrets.GITHUB_TOKEN }}
-          script: |
-            const forceAll = process.env.FORCE_ALL === 'true';
-            const maxPrs = Math.min(Number(process.env.MAX_PRS) || 8, 20);
-            const marker = '<!-- continuous-agent-ops -->';
-            const debounceMs = 90 * 60 * 1000; // 90 minutes
-            const staleMs = 3 * 60 * 60 * 1000; // 3 hours without agent activity
-
-            const agentLogins = [
-              'google-labs-jules', 'devin-ai-integration', 'coderabbitai',
-              'github-actions', 'copilot', 'gitar-bot', 'blocksorg',
-            ];
-            const isAgent = (login) => {
-              const l = (login || '').toLowerCase();
-              return agentLogins.some(a => l.includes(a));
-            };
-
-            // List open non-draft PRs (newest activity first)
-            const prs = await github.paginate(github.rest.pulls.list, {
-              owner: context.repo.owner,
-              repo: context.repo.repo,
-              state: 'open',
-              sort: 'updated',
-              direction: 'desc',
-              per_page: 50,
-            });
-
-            const eligible = prs.filter(p => !p.draft && !p.head.repo?.fork);
-            core.info(`Open non-draft PRs: ${eligible.length}`);
-
-            let acted = 0;
-            const report = [];
-
-            for (const pr of eligible) {
-              if (acted >= maxPrs) break;
-
-              // Fresh mergeability
-              let mergeable = pr.mergeable;
-              let mergeableState = pr.mergeable_state;
-              try {
-                const { data: fresh } = await github.rest.pulls.get({
-                  owner: context.repo.owner,
-                  repo: context.repo.repo,
-                  pull_number: pr.number,
-                });
-                mergeable = fresh.mergeable;
-                mergeableState = fresh.mergeable_state;
-              } catch (e) {
-                core.info(`PR #${pr.number} get skip: ${e.message}`);
-              }
-
-              // Comments for debounce + last agent activity
-              const comments = await github.paginate(github.rest.issues.listComments, {
-                owner: context.repo.owner,
-                repo: context.repo.repo,
-                issue_number: pr.number,
-                per_page: 100,
-              });
-
-              const lastOps = comments
-                .filter(c => c.body && c.body.includes(marker))
-                .sort((a, b) => new Date(b.created_at) - new Date(a.created_at))[0];
-              if (lastOps && !forceAll) {
-                const age = Date.now() - new Date(lastOps.created_at).getTime();
-                if (age < debounceMs) {
-                  core.info(`PR #${pr.number}: debounced (${Math.round(age/60000)}m ago)`);
-                  continue;
-                }
-              }
-
-              const lastAgentComment = comments
-                .filter(c => isAgent(c.user?.login))
-                .sort((a, b) => new Date(b.created_at) - new Date(a.created_at))[0];
-              const lastAgentAge = lastAgentComment
-                ? Date.now() - new Date(lastAgentComment.created_at).getTime()
-                : Infinity;
-
-              // Open review threads (unresolved)
-              let openThreads = 0;
-              try {
-                const q = `
-                  query($owner:String!, $repo:String!, $number:Int!) {
-                    repository(owner:$owner, name:$repo) {
-                      pullRequest(number:$number) {
-                        reviewThreads(first: 50) {
-                          nodes { isResolved }
-                        }
-                      }
-                    }
-                  }`;
-                const result = await github.graphql(q, {
-                  owner: context.repo.owner,
-                  repo: context.repo.repo,
-                  number: pr.number,
-                });
-                const nodes = result?.repository?.pullRequest?.reviewThreads?.nodes || [];
-                openThreads = nodes.filter(n => !n.isResolved).length;
-              } catch (e) {
-                core.info(`PR #${pr.number} threads skip: ${e.message}`);
-              }
-
-              const dirty = mergeableState === 'dirty' || mergeable === false;
-              const blocked = mergeableState === 'blocked' || mergeableState === 'unstable';
-              const stale = lastAgentAge > staleMs;
-              const needsWork = dirty || (blocked && openThreads > 0) || (stale && openThreads > 0);
-
-              if (!needsWork && !forceAll) {
-                core.info(`PR #${pr.number}: clean enough (state=${mergeableState} threads=${openThreads})`);
-                continue;
-              }
-
-              // Build action prompt
-              const reasons = [];
-              if (dirty) reasons.push('merge conflict / dirty vs base');
-              if (blocked && openThreads > 0) reasons.push(`${openThreads} unresolved review thread(s)`);
-              if (stale) reasons.push(`stale agent activity (${Math.round(lastAgentAge/3600000)}h)`);
-              if (forceAll) reasons.push('forced sweep');
-
-              const instructions = [];
-              if (dirty) {
-                instructions.push('1. Rebase or merge base into head; resolve conflicts; push.');
-                instructions.push('2. Prefer content-preserving resolution; cite Implements: if board item applies.');
-              }
-              if (openThreads > 0) {
-                instructions.push(`${dirty ? '3' : '1'}. Address **all open review threads** (CodeRabbit, Devin, Copilot).`);
-                instructions.push(`${dirty ? '4' : '2'}. Prefer minimal diffs; preserve Sentinel 0o600/0o700.`);
-              }
-              instructions.push('Push commits to the existing head branch. Do not retarget base without cause.');
-              instructions.push('Skip pure nits only if they conflict with security/gates.');
-              instructions.push('If the PR is superseded, close it with a pointer to the replacement.');
-
-              const body = [
-                marker,
-                `sha: ${pr.head.sha}`,
-                `state: ${mergeableState}`,
-                `threads_open: ${openThreads}`,
-                '',
-                `@jules **Continuous ops** (GHA agent-continuous-ops) — unattended advance.`,
-                '',
-                `PR #${pr.number} · \`${pr.head.ref}\` → \`${pr.base.ref}\``,
-                `**Why:** ${reasons.join('; ')}`,
-                '',
-                '### Instructions',
-                ...instructions,
-                '',
-                'Read AGENTS.md. No Class 3/4 artifacts. No secret leaks.',
-                'Agent: Grok orchestration · Profile: https://x.com/grok',
-              ].join('\n');
-
-              await github.rest.issues.createComment({
-                owner: context.repo.owner,
-                repo: context.repo.repo,
-                issue_number: pr.number,
-                body,
-              });
-
-              acted += 1;
-              report.push(`#${pr.number} (${reasons.join(', ')})`);
-              core.info(`PR #${pr.number}: nudged — ${reasons.join('; ')}`);
-
-              // Brief pause to avoid secondary rate limits
-              await new Promise(r => setTimeout(r, 1500));
-            }
-
-            core.summary
-              .addHeading('Continuous agent ops')
-              .addRaw(`Acted on **${acted}** PR(s) this sweep.`)
-              .addList(report.length ? report : ['(none eligible)'])
-              .write();
-
-            core.setOutput('acted', String(acted));
diff --git a/.github/workflows/deepseek-ci.yml b/.github/workflows/deepseek-ci.yml
new file mode 100644
index 000000000..1ae46f2eb
--- /dev/null
+++ b/.github/workflows/deepseek-ci.yml
@@ -0,0 +1,72 @@
+name: DeepSeek CI – Agentic Automation
+
+on:
+  pull_request:
+    types: [opened, synchronize, reopened]
+  workflow_dispatch:
+    inputs:
+      event_payload:
+        description: 'JSON event payload (optional)'
+        required: false
+        default: '{}'
+
+concurrency:
+  group: deepseek-ci-${{ github.event.pull_request.number || github.run_id }}
+  cancel-in-progress: true
+
+permissions: {}
+
+jobs:
+  deepseek-agent:
+    runs-on: ubuntu-latest
+    permissions:
+      contents: write
+      pull-requests: write
+      issues: write
+      actions: read
+
+    env:
+      OPERATOR_TOKEN: ${{ secrets.OPERATOR_TOKEN }}
+      GITHUB_EVENT: ${{ toJson(github.event) }}
+      DEEPSEEK_WASM_PATH: ./deepcli/deepseek.wasm
+
+    steps:
+      - name: Debounce rapid-succession commits
+        run: sleep 10
+
+      - name: Checkout repository
+        uses: actions/checkout@08c6903cd8c0fde910a37f88322edcfb5dd907a8 # v5.0.0
+        with:
+          fetch-depth: 0
+          token: ${{ secrets.OPERATOR_TOKEN || secrets.GITHUB_TOKEN }}
+
+      - name: Setup Node.js (for PoW WASM)
+        uses: actions/setup-node@0a44ba7841725637a19e28fa30b79a866c81b0a6 # v4.0.4
+        with:
+          node-version: '20'
+
+      - name: Setup Python
+        uses: actions/setup-python@39cd14951b08e74b54015e9e001cdefcf80e669f # v5.1.1
+        with:
+          python-version: '3.11'
+          cache: 'pip'
+
+      - name: Install Python dependencies
+        run: |
+          pip install curl_cffi requests pyyaml click
+
+      - name: Run DeepSeek CI Agent
+        env:
+          PYTHONPATH: ${{ github.workspace }}/multi-ai-cli
+        run: |
+          python -m ci_mode \
+            --workspace "$GITHUB_WORKSPACE" \
+            --cache-dir "${{ runner.temp }}/deepseek-cache"
+
+      - name: Upload results (if any)
+        uses: actions/upload-artifact@b4b15b8c7c6ac21ea08fcf65892d2ee8f75cf882 # v4.4.3
+        if: always()
+        with:
+          name: deepseek-output
+          path: ./deepseek_output.json
+          if-no-files-found: ignore
diff --git a/.gitignore b/.gitignore
index ad7874842..66edcc5cc 100644
--- a/.gitignore
+++ b/.gitignore
@@ -36,6 +36,9 @@ pipeline_log.jsonl
 *.db
 *.sqlite
 *.sqlite3
+
+# === DeepSeek v4-Pro CI ===
+deepseek_output.json
 *.tar
 *.gz
 *.bz2
diff --git a/docs/ops/session-context-management.md b/docs/ops/session-context-management.md
deleted file mode 100644
index a8f6746e3..000000000
--- a/docs/ops/session-context-management.md
+++ /dev/null
@@ -1,87 +0,0 @@
-# Session && Context Management (GitHub Actions)
-
-**Issue:** [#118](https://github.com/timerloggedout-spec/termux-monorepo/issues/118)  
-**Related:** #109 · #112 (merged) · #114 · #72 · PR #3 (session-store untrack)
-
-## Why “cached data was removed”
-
-| Layer | What it holds | Persistence policy |
-|-------|---------------|--------------------|
-| **Auth session** | Cookies, Bearer tokens, PoW answers/signatures, browser profile | **Ephemeral only** — `$RUNNER_TEMP` / `$HOME` override inside the job; discarded at job end. Never Actions cache, never commit, never artifact. |
-| **Work context** | Task key, last-reviewed SHA, provider `session_id` *hash* (not the token), last action type, short status flags | **Durable (non-secret)** — Actions cache + optional PR/issue comment marker. |
-
-PR **#112 / #114** deliberately dropped DeepSeek session/cookie caching after CodeRabbit + Devin required that no session material land in the tree or cross-run secrets surface. That fixed a security class problem; it also removed the accidental ability to “pick up the same chat.”
-
-#118 is the product follow-up: **continue the *work*, not the *credentials*.**
-
-## What you can continue / regenerate / edit
-
-Using `.github/actions/agent-context-store`:
-
-1. **Load** prior context for a stable key (`pr-123-review`, `issue-118-triage`, …).
-2. **Decide** whether to skip (same SHA already reviewed), re-diff only changed files, or regenerate.
-3. **Save** updated non-secret fields after the run.
-4. Optional **comment marker** `<!-- agent-ctx:<key> -->` so humans/agents see a pointer on the issue/PR.
-
-Forbidden keys are stripped on both load and save: `token`, `cookie(s)`, `authorization`, `api_key`, `pow_*`, `password`, `secret`, and any key containing those substrings.
-
-## Usage sketch
-
-```yaml
-- name: Load work context
-  id: ctx
-  uses: ./.github/actions/agent-context-store
-  with:
-    mode: load
-    context_key: pr-${{ github.event.pull_request.number }}-review
-
-- name: Agent step
-  # use fromJSON(steps.ctx.outputs.context_json).last_sha etc.
-
-- name: Save work context
-  if: always()
-  uses: ./.github/actions/agent-context-store
-  with:
-    mode: save
-    context_key: pr-${{ github.event.pull_request.number }}-review
-    context_json: |
-      {"last_sha":"${{ github.event.pull_request.head.sha }}","last_action":"review","provider_session_hash":"${{ steps.hash.outputs.value }}"}
-    github_token: ${{ secrets.GITHUB_TOKEN }}
-    target_number: ${{ github.event.pull_request.number }}
-    repository: ${{ github.repository }}
-```
-
-## Auth session path (DeepSeek web-wrapper and similar)
-
-Keep the #114 pattern:
-
-```bash
-export HOME="${RUNNER_TEMP}/deepseek-webwrapper-home"
-mkdir -p "$HOME"
-# … create_session / chat …
-rm -rf "${RUNNER_TEMP}/deepseek-webwrapper-home"   # always()
-```
-
-If a provider returns a `session_id`, store only a **hash** (or the last 8 chars for logs) inside work context — never the full id + token pair.
-
-## Gemini / peer review continuity
-
-Existing Gemini review already uses soft-budget skip comments and router peers. Prefer embedding **SHA bookmarks** in work context rather than re-introducing cookie caches.
-
-## Operator checklist
-
-- [ ] Secrets stay in GitHub Secrets / local `~/.deepcli` (ignored).
-- [ ] No `session.json` / `cookies_*.json` in the index (see `.gitignore` + #3).
-- [ ] New workflows call `agent-context-store` for any multi-run agent task.
-- [ ] Regenerating a response = new invoke with prior `last_sha` / task flags loaded, not replaying auth cookies.
-
-## Trace
-
-| Artifact | Role |
-|----------|------|
-| #118 | Continuity requirement |
-| #109 | DeepSeek CI integration source |
-| #112 | Executable workflow + ephemeral session policy (merged) |
-| #114 | Peer routing + web-wrapper opt-in (open) |
-| #72 | Quota + session continuation foundations |
-| #3 | Untrack historical session stores |
diff --git a/docs/proposals/active/rate-limit-rotation/ITEMS.md b/docs/proposals/active/rate-limit-rotation/ITEMS.md
index 3affb6d14..97a6be93e 100644
--- a/docs/proposals/active/rate-limit-rotation/ITEMS.md
+++ b/docs/proposals/active/rate-limit-rotation/ITEMS.md
@@ -6,9 +6,9 @@
 | RL-02 | model-router composite action | P0 | grok | done | .github/actions/model-router |
 | RL-03 | Wire triage → Flash-Lite | P0 | grok | done | gemini-triage.yml |
 | RL-04 | Wire review → Flash w/ Lite fallback | P0 | grok | done | gemini-review.yml |
-| RL-05 | OpenRouter fallback path | P0 | grok | executing | #123 landed ELO+poll router; http-llm-invoke fallback still tracked |
+| RL-05 | OpenRouter fallback path | P0 | grok | foundation | skip=true when Gemini exhausted; real HTTP invoke still TODO |
 | RL-06 | Tighten job-gate daily-limit (900→100) | P1 | grok | todo | pair with #81 |
-| RL-07 | Prompt compression for agent workflows | P1 | | todo | issue #90; PR #126 |
+| RL-07 | Prompt compression for agent workflows | P1 | | todo | issue #90 |
 | RL-08 | OmniRoute hub integration surface | P2 | | todo | issue #91 |
 | RL-09 | Per-model Linear quota dashboard | P2 | | todo | |
 | RL-10 | Merge/rebase #81 quota-gate onto master | P0 | | todo | PR #81 |
@@ -18,5 +18,4 @@
 | RL-14 | Devin Apply Suggestions automation | P1 | | partial | enable Auto-Fix in Devin Settings; no public click-API |
 | RL-15 | Honest OpenRouter (no false route) | P0 | grok | done | model-router skip=true always when Gemini exhausted |
 | RL-16 | Shared global counter (not per-branch cache) | P2 | | todo | gist/issue optimistic concurrency |
-| RL-17 | Model availability polling & ELO (3L0) routing | P0 | jules | done | #123 scripts/model_router.py, model-success-matrix.yaml |
-| RL-18 | DeepSeek CI peer path (no Class 3/4 cache) | P0 | jules | blocked | #134 security hold — ephemeral session only |
+| RL-17 | Model availability polling & ELO (3L0) routing | P0 | jules | done | scripts/model_router.py, docs/schemas/model-success-matrix.yaml |
diff --git a/multi-ai-cli/ci_mode.py b/multi-ai-cli/ci_mode.py
new file mode 100644
index 000000000..d0b3f4169
--- /dev/null
+++ b/multi-ai-cli/ci_mode.py
@@ -0,0 +1,138 @@
+#!/usr/bin/env python3
+"""
+CI mode for multi-ai-cli – runs the agent non-interactively using DeepSeek,
+reading GitHub event payload, and outputting JSON securely.
+"""
+import os
+import sys
+import json
+import argparse
+import subprocess
+from pathlib import Path
+
+# Use relative module paths inside multi-ai-cli
+sys.path.insert(0, str(Path(__file__).parent))
+
+from core.session_manager import SessionManager
+import backends.deepseek as ds_mod
+from backends.deepseek import DeepSeekBackend
+
+def run_ci(event, workspace, operator_token):
+    """
+    Non-interactive agent loop using DeepSeek web-wrapper.
+    """
+    # Dynamically resolve WASM_SOLVER path in GHA
+    ds_mod.WASM_SOLVER = Path(workspace) / "deepcli" / "pow_solver.js"
+
+    # Set up GitHub CLI env with token
+    gh_env = os.environ.copy()
+    if operator_token:
+        gh_env['GH_TOKEN'] = operator_token
+
+    pr_number = event.get('pull_request', {}).get('number')
+    repo = event.get('repository', {}).get('full_name') or os.environ.get('GITHUB_REPOSITORY')
+    action = event.get('action')
+
+    decisions = []
+
+    if action in ['opened', 'synchronize', 'reopened'] and pr_number and repo:
+        # Get PR diff
+        diff_cmd = ['gh', 'pr', 'diff', str(pr_number), '--repo', repo]
+        try:
+            diff = subprocess.check_output(diff_cmd, env=gh_env, text=True)
+        except Exception as e:
+            diff = f"Could not retrieve diff: {e}"
+
+        # Initialize the DeepSeek backend and perform code review
+        try:
+            # DeepSeek token can be loaded from env or config.yaml
+            if not os.environ.get("DEEPSEEK_TOKEN"):
+                os.environ["DEEPSEEK_TOKEN"] = operator_token or ""
+
+            mgr = SessionManager()
+            backend = DeepSeekBackend(mgr)
+
+            prompt = f"You are a code reviewer. Analyze the diff and suggest improvements:\n\n{diff[:8000]}"
+            analysis = backend.send_message(prompt, [])
+
+            # Ensure we only post PR comment on successful analysis (skip error/mock strings)
+            if analysis and not analysis.startswith("Error:") and not analysis.startswith("[No content returned]"):
+                # Tracing signature metadata suffix
+                signature = f"\n\n---\n*Bot Review powered by @deepseek-cli{{provider: deepseek, model: deepseek-reasoner}}*"
+                comment_body = analysis[:1900] + signature
+
+                # Comment on the PR
+                comment_cmd = ['gh', 'pr', 'comment', str(pr_number), '--body', comment_body, '--repo', repo]
+                try:
+                    subprocess.run(comment_cmd, env=gh_env, check=False)
+                except Exception as e:
+                    print(f"Failed to post PR comment: {e}")
+
+                decisions.append({
+                    'type': 'pr_review',
+                    'pr': pr_number,
+                    'summary': comment_body[:200],
+                })
+            else:
+                print(f"Warning: DeepSeek analysis returned empty or invalid review: {analysis}", file=sys.stderr)
+        except Exception as e:
+            print(f"Error during DeepSeek analysis: {e}", file=sys.stderr)
+
+    return {
+        'actions': decisions,
+        'event': event.get('action'),
+        'pr': pr_number,
+        'provider_used': 'deepseek',
+    }
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--workspace', required=True, help='GitHub workspace path')
+    parser.add_argument('--cache-dir', default='/tmp/deepseek-cache', help='Directory for session cache')
+    args = parser.parse_args()
+
+    # Enforce secure directory creation (0o700)
+    os.makedirs(args.cache_dir, exist_ok=True)
+    try:
+        os.chmod(args.cache_dir, 0o700)
+    except Exception:
+        pass
+
+    # Read event from GITHUB_EVENT_PATH
+    event = {}
+    event_path = os.environ.get('GITHUB_EVENT_PATH')
+    if event_path and os.path.exists(event_path):
+        try:
+            with open(event_path, 'r') as f:
+                event = json.load(f)
+        except Exception as e:
+            print(f"Warning: Failed to load event payload from GITHUB_EVENT_PATH: {e}")
+
+    # Run DeepSeek CI
+    result = run_ci(
+        event=event,
+        workspace=args.workspace,
+        operator_token=os.environ.get('OPERATOR_TOKEN')
+    )
+
+    # Secure output serialization with 0o600 permissions
+    output_path = Path('deepseek_output.json')
+    flags = os.O_WRONLY | os.O_CREAT | os.O_TRUNC
+    mode = 0o600
+    try:
+        fd = os.open(output_path, flags, mode)
+        with open(fd, 'w') as f:
+            json.dump(result, f, indent=2)
+    except Exception:
+        with open(output_path, 'w') as f:
+            json.dump(result, f, indent=2)
+
+    try:
+        os.chmod(output_path, 0o600)
+    except Exception:
+        pass
+
+    print(f"✅ CI run completed. Decisions: {result.get('actions', [])}")
+
+if __name__ == '__main__':
+    main()
diff --git a/scripts/model_router.py b/scripts/model_router.py
old mode 100644
new mode 100755
index bd464450a..4bd008905
--- a/scripts/model_router.py
+++ b/scripts/model_router.py
@@ -4,12 +4,7 @@
 Implements Model Availability Polling with smart temporal caching,
 dynamic ELO-based (3L0) ranking, and soft-budget validation.
 
-Soft budgets raised 2026-08-10 (OPERATOR) to exceed prior free-tier headroom.
-See .github/connectors/llm-peers.yaml for authorized catalog.
-
-LEGACY_MODELS is the conservative catalog-unavailable allow-list: only models
-proven in production. Newly listed free models must appear in a live/stale
-OpenRouter poll before they are selected.
+This script replaces rigid bash logic with robust Python standard library code.
 """
 
 import os
@@ -17,39 +12,49 @@
 import re
 import json
 import urllib.request
+import ssl
 import time
 
 COUNTER_DIR = os.environ.get("COUNTER_DIR", "/tmp/model-router")
 
-# Proven-only fallback when OpenRouter catalog cannot be fetched.
-# Do NOT add unverified free models here — they belong in role_peers and the live poll.
+# Previously known-good models to fall back on when no live catalog is available
 LEGACY_MODELS = {
     "meta-llama/llama-3.3-70b-instruct:free",
     "google/gemma-3-12b-it:free",
     "qwen/qwen3-coder:free",
     "deepseek/deepseek-r1:free",
-    "auto/best-free",
+    "auto/best-free"
 }
 
 def parse_yaml(filepath):
+    """
+    Minimal robust YAML parser for standard key-value and indentation structures.
+    Does not require external dependencies like PyYAML.
+    """
     if not os.path.exists(filepath):
         return {}
+
     result = {}
     path = []
+
     with open(filepath, 'r', encoding='utf-8') as f:
         for line in f:
             stripped = line.strip()
             if not stripped or stripped.startswith('#'):
                 continue
+
             indent = len(line) - len(line.lstrip())
+
             while path and path[-1][0] >= indent:
                 path.pop()
+
             if not path:
                 parent = result
                 parent_key = None
             else:
                 parent = path[-1][1]
                 parent_key = path[-1][2]
+
             if stripped.startswith('- '):
                 item_str = stripped[2:].strip()
                 if ':' in item_str and not (item_str.startswith('"') or item_str.startswith("'")):
@@ -69,6 +74,7 @@ def parse_yaml(filepath):
                         else: item_val = int(item_val)
                     except ValueError:
                         pass
+
                 if isinstance(parent, dict) and parent_key is not None:
                     if parent_key not in parent or not isinstance(parent[parent_key], list):
                         parent[parent_key] = []
@@ -77,14 +83,17 @@ def parse_yaml(filepath):
                         sub_key = list(item_val.keys())[0]
                         path.append((indent, item_val, sub_key))
                 continue
+
             m = re.match(r'^("[^"]+"|\'[^\']+\'|[^:]+):\s*(.*)$', stripped)
             if not m:
                 continue
+
             key = m.group(1).strip().strip('"\'')
             val = m.group(2).strip()
             if '#' in val:
                 val = val.split('#')[0].strip()
             val = val.strip('"\'')
+
             if not val:
                 new_dict = {}
                 if isinstance(parent, dict):
@@ -106,18 +115,25 @@ def parse_yaml(filepath):
                                 parsed_val = int(val)
                         except ValueError:
                             parsed_val = val
+
                 if isinstance(parent, dict):
                     parent[key] = parsed_val
                 path.append((indent, parent, key))
+
     return result
 
+
 def fetch_openrouter_free_models():
-    """Poll OpenRouter models endpoint; return free model ids or None on failure."""
+    """
+    Polls the OpenRouter models endpoint to find available free models.
+    """
     url = "https://openrouter.ai/api/v1/models"
     try:
         req = urllib.request.Request(url, headers={"User-Agent": "Mozilla/5.0"})
+        # Use standard SSL verification by default for secure transfers
         with urllib.request.urlopen(req, timeout=5) as response:
             data = json.loads(response.read().decode('utf-8'))
+
             free_models = []
             for m in data.get('data', []):
                 m_id = m.get('id', '')
@@ -126,6 +142,7 @@ def fetch_openrouter_free_models():
                 if m_id.endswith(':free'):
                     is_free = True
                 else:
+                    # Robust check converting pricing prompt/completion to float
                     try:
                         p_prompt = float(pricing.get('prompt', 0.0))
                         p_compl = float(pricing.get('completion', 0.0))
@@ -133,8 +150,11 @@ def fetch_openrouter_free_models():
                             is_free = True
                     except (ValueError, TypeError):
                         pass
+
                 if is_free and m_id:
                     free_models.append(m_id)
+
+            # Prevent caching an empty list if there's a temporary API anomaly
             if not free_models:
                 return None
             return free_models
@@ -142,11 +162,17 @@ def fetch_openrouter_free_models():
         sys.stderr.write(f"Warning: Failed to poll OpenRouter models ({e}). Using hardcoded fallback list.\n")
         return None
 
+
 def fetch_openrouter_free_models_cached():
-    """1h cache; on miss return stale cache if present, else live poll."""
+    """
+    Retrieves the list of OpenRouter free models using a 1-hour temporal file cache.
+    Optimizes polling intervals to avoid redundant network overhead.
+    """
     cache_file = os.path.join(COUNTER_DIR, "openrouter_models_cache.json")
     cached_models = None
     cached_time = 0
+
+    # Try loading existing cache
     if os.path.exists(cache_file):
         try:
             with open(cache_file, "r") as f:
@@ -155,24 +181,38 @@ def fetch_openrouter_free_models_cached():
                 cached_time = cache_data.get("timestamp", 0)
         except Exception:
             pass
+
+    # If cache is valid (less than 1 hour old), return it immediately
     if cached_models is not None and (time.time() - cached_time < 3600):
         sys.stderr.write("Using cached OpenRouter free models list.\n")
         return cached_models
+
+    # Cache is missing or expired, fetch fresh list from API
     models = fetch_openrouter_free_models()
     if models is not None:
         try:
             os.makedirs(COUNTER_DIR, exist_ok=True)
             with open(cache_file, "w") as f:
-                json.dump({"timestamp": time.time(), "models": models}, f)
+                json.dump({
+                    "timestamp": time.time(),
+                    "models": models
+                }, f)
         except Exception:
             pass
         return models
+
+    # Fetch failed! Fall back to stale cached models if available to avoid unverified routes
     if cached_models is not None:
         sys.stderr.write("Warning: OpenRouter fresh poll failed. Falling back to stale cached models list.\n")
         return cached_models
+
     return None
 
+
 def get_usage(provider, model):
+    """
+    Returns the daily counter for a model.
+    """
     key = f"{provider}/{model}" if provider != "gemini" else model
     filename = os.path.join(COUNTER_DIR, key.replace('/', '_') + ".txt")
     if os.path.exists(filename):
@@ -183,7 +223,11 @@ def get_usage(provider, model):
             return 0
     return 0
 
+
 def increment_usage(provider, model):
+    """
+    Increments and saves the daily counter for a model.
+    """
     key = f"{provider}/{model}" if provider != "gemini" else model
     os.makedirs(COUNTER_DIR, exist_ok=True)
     filename = os.path.join(COUNTER_DIR, key.replace('/', '_') + ".txt")
@@ -192,38 +236,43 @@ def increment_usage(provider, model):
         f.write(str(current + 1))
     return current + 1
 
+
 def main():
+    # Load input arguments or environment variables
     role = os.environ.get("ROLE", "triage")
     has_omni = os.environ.get("HAS_OMNI", "false").lower() == "true"
     has_openrouter = os.environ.get("HAS_OPENROUTER", "false").lower() == "true"
     has_gemini = os.environ.get("HAS_GEMINI", "true").lower() == "true"
 
+    # Load configuration schemas
     matrix_path = "docs/schemas/model-success-matrix.yaml"
     success_matrix = parse_yaml(matrix_path)
 
+    # 1. Poll OpenRouter models for availability (only if has_openrouter is enabled)
     polled_free_models = None
     if has_openrouter:
         polled_free_models = fetch_openrouter_free_models_cached()
 
-    # ELEVATED soft limits (OPERATOR 2026-08-10) — exceed prior capacity.
-    # Soft gates only; downstream 429 must still re-route / fall through.
+    # 2. Define candidates per role and retrieve limit mapping
+    # Defaults in case schemas are missing or incomplete
     limits = {
-        "omni/auto/best-free": {"triage": 400, "review": 250, "invoke": 400},
-        "openrouter/meta-llama/llama-3.3-70b-instruct:free": {"triage": 80, "review": 80, "invoke": 80},
-        "openrouter/google/gemma-3-12b-it:free": {"triage": 80, "review": 80, "invoke": 80},
-        "openrouter/qwen/qwen3-coder:free": {"triage": 60, "review": 60, "invoke": 60},
-        "openrouter/deepseek/deepseek-r1:free": {"triage": 40, "review": 40, "invoke": 40},
-        "openrouter/google/gemma-4-31b-it:free": {"triage": 80, "review": 80, "invoke": 80},
-        "openrouter/google/gemma-4-26b-a4b-it:free": {"triage": 80, "review": 80, "invoke": 80},
-        "openrouter/cohere/north-mini-code:free": {"triage": 60, "review": 60, "invoke": 60},
+        "omni/auto/best-free": {"triage": 200, "review": 120, "invoke": 200},
+        "openrouter/meta-llama/llama-3.3-70b-instruct:free": {"triage": 40, "review": 40, "invoke": 40},
+        "openrouter/google/gemma-3-12b-it:free": {"triage": 40, "review": 40, "invoke": 40},
+        "openrouter/qwen/qwen3-coder:free": {"triage": 30, "review": 30, "invoke": 30},
+        "openrouter/deepseek/deepseek-r1:free": {"triage": 20, "review": 20, "invoke": 20},
+        "openrouter/google/gemma-4-31b-it:free": {"triage": 40, "review": 40, "invoke": 40},
+        "openrouter/google/gemma-4-26b-a4b-it:free": {"triage": 40, "review": 40, "invoke": 40},
+        "openrouter/cohere/north-mini-code:free": {"triage": 30, "review": 30, "invoke": 30},
         "gemini-3.1-flash-lite": {"triage": 450, "review": 450, "invoke": 450},
         "gemini-3.5-flash-lite": {"triage": 450, "review": 450, "invoke": 450},
-        "gemini-2.5-flash-lite": {"triage": 20, "review": 20, "invoke": 20},
-        "gemini-3.5-flash": {"triage": 20, "review": 20, "invoke": 20},
-        "gemini-2.5-flash": {"triage": 20, "review": 20, "invoke": 20},
-        "gemini-3-flash": {"triage": 20, "review": 20, "invoke": 20},
+        "gemini-2.5-flash-lite": {"triage": 15, "review": 15, "invoke": 15},
+        "gemini-3.5-flash": {"triage": 15, "review": 15, "invoke": 15},
+        "gemini-2.5-flash": {"triage": 15, "review": 15, "invoke": 15},
+        "gemini-3-flash": {"triage": 15, "review": 15, "invoke": 15},
     }
 
+    # Preferred lists per role
     role_peers = {
         "triage": [
             ("omni", "auto/best-free"),
@@ -255,51 +304,73 @@ def main():
         "invoke": ["gemini-3.1-flash-lite", "gemini-3.5-flash-lite"],
     }
 
+    # 3. Score and rank peer candidates
     peer_candidates = []
     for provider, model in role_peers.get(role, []):
         if provider == "omni" and not has_omni:
             continue
         if provider == "openrouter" and not has_openrouter:
             continue
+
+        # Security Assertion: guard against non-free model ids on OpenRouter paths
         if provider == "openrouter" and not model.endswith(":free"):
-            sys.stderr.write(f"Warning: non-free model ID '{model}' skipped.\n")
+            sys.stderr.write(f"Warning: Model router security check failed — non-free model ID '{model}' skipped.\n")
             continue
+
+        # Cross-reference OpenRouter polled availability
         if provider == "openrouter":
             if polled_free_models is not None:
                 if model not in polled_free_models:
-                    sys.stderr.write(f"Warning: OpenRouter model {model} not in polled catalog. Skipping.\n")
+                    sys.stderr.write(f"Warning: OpenRouter model {model} is not currently available/free in polled catalog. Skipping.\n")
                     continue
             else:
+                # No catalog (fresh or cached) available at all! Skip unverified new models conservatively.
                 if model not in LEGACY_MODELS:
-                    sys.stderr.write(f"Warning: No catalog. Skipping unverified '{model}'.\n")
+                    sys.stderr.write(f"Warning: No OpenRouter catalog available. Skipping unverified new model '{model}' conservatively.\n")
                     continue
+
+        # Calculate ELO score from success matrix (ELO ≈ 3L0)
         model_entry = success_matrix.get("models", {}).get(model, {})
         elo = model_entry.get("elo", 1000)
         suitability = model_entry.get("role_suitability", {}).get(role, 1.0)
         score = elo * suitability
-        peer_candidates.append({"provider": provider, "model": model, "score": score})
 
+        peer_candidates.append({
+            "provider": provider,
+            "model": model,
+            "score": score
+        })
+
+    # Sort peers by ELO score descending
     peer_candidates.sort(key=lambda x: x["score"], reverse=True)
 
+    # 4. Attempt routing to highest ranked peer with capacity
     for candidate in peer_candidates:
         prov = candidate["provider"]
         mod = candidate["model"]
         limit_dict = limits.get(f"{prov}/{mod}", {})
         limit = limit_dict.get(role, limit_dict.get("default", 40))
+
         used = get_usage(prov, mod)
         if used < limit:
             new_used = increment_usage(prov, mod)
-            reason = f"role={role} peer={prov} model={mod} used={new_used}/{limit} (ranked score={candidate['score']})"
             print(f"::set-output name=provider::{prov}")
             print(f"::set-output name=model::{mod}")
             print(f"::set-output name=skip::false")
-            print(f"::set-output name=reason::{reason}")
+            print(f"::set-output name=reason::role={role} peer={prov} model={mod} used={new_used}/{limit} (ranked score={candidate['score']})")
+
+            # GITHUB_OUTPUT file support (newer GitHub Actions convention)
             if "GITHUB_OUTPUT" in os.environ:
                 with open(os.environ["GITHUB_OUTPUT"], "a") as go:
-                    go.write(f"provider={prov}\nmodel={mod}\nskip=false\nreason={reason}\n")
+                    go.write(f"provider={prov}\n")
+                    go.write(f"model={mod}\n")
+                    go.write(f"skip=false\n")
+                    go.write(f"reason=role={role} peer={prov} model={mod} used={new_used}/{limit} (ranked score={candidate['score']})\n")
             return
+
         sys.stderr.write(f"Peer soft budget exhausted: {prov}/{mod} ({used}/{limit})\n")
 
+    # 5. Fallback to Gemini residuals if peers are exhausted
     if has_gemini:
         gemini_candidates = []
         for model in role_residuals.get(role, []):
@@ -307,47 +378,70 @@ def main():
             elo = model_entry.get("elo", 1000)
             suitability = model_entry.get("role_suitability", {}).get(role, 1.0)
             score = elo * suitability
-            gemini_candidates.append({"provider": "gemini", "model": model, "score": score})
+
+            gemini_candidates.append({
+                "provider": "gemini",
+                "model": model,
+                "score": score
+            })
+
+        # Sort Gemini candidates by ELO score descending
         gemini_candidates.sort(key=lambda x: x["score"], reverse=True)
+
         for candidate in gemini_candidates:
             mod = candidate["model"]
             limit_dict = limits.get(mod, {})
             limit = limit_dict.get(role, limit_dict.get("default", 15))
+
             used = get_usage("gemini", mod)
             if used < limit:
                 new_used = increment_usage("gemini", mod)
-                reason = f"role={role} gemini={mod} used={new_used}/{limit} (ranked score={candidate['score']}) residual"
                 print(f"::set-output name=provider::gemini")
                 print(f"::set-output name=model::{mod}")
                 print(f"::set-output name=skip::false")
-                print(f"::set-output name=reason::{reason}")
+                print(f"::set-output name=reason::role={role} gemini={mod} used={new_used}/{limit} (ranked score={candidate['score']}) residual")
+
                 if "GITHUB_OUTPUT" in os.environ:
                     with open(os.environ["GITHUB_OUTPUT"], "a") as go:
-                        go.write(f"provider=gemini\nmodel={mod}\nskip=false\nreason={reason}\n")
+                        go.write(f"provider=gemini\n")
+                        go.write(f"model={mod}\n")
+                        go.write(f"skip=false\n")
+                        go.write(f"reason=role={role} gemini={mod} used={new_used}/{limit} (ranked score={candidate['score']}) residual\n")
                 return
+
             sys.stderr.write(f"Gemini residual soft budget exhausted: {mod} ({used}/{limit})\n")
 
-    reason = f"all free paths exhausted (omni/openrouter peers + gemini residual) role={role}"
-    print("::set-output name=provider::none")
-    print("::set-output name=model::")
-    print("::set-output name=skip::true")
-    print(f"::set-output name=reason::{reason}")
+    # 6. All free options exhausted
+    print(f"::set-output name=provider::none")
+    print(f"::set-output name=model::")
+    print(f"::set-output name=skip::true")
+    print(f"::set-output name=reason::all free paths exhausted (omni/openrouter peers + gemini residual) role={role}")
+
     if "GITHUB_OUTPUT" in os.environ:
         with open(os.environ["GITHUB_OUTPUT"], "a") as go:
-            go.write(f"provider=none\nmodel=\nskip=true\nreason={reason}\n")
+            go.write(f"provider=none\n")
+            go.write(f"model=\n")
+            go.write(f"skip=true\n")
+            go.write(f"reason=all free paths exhausted (omni/openrouter peers + gemini residual) role={role}\n")
+
 
 if __name__ == "__main__":
     try:
         main()
     except Exception as e:
-        sys.stderr.write(f"Error: Model Router crashed: {e}\n")
+        sys.stderr.write(f"Error: Model Router crashed during execution: {e}\n")
+        # Graceful degradation fallback outputs
         print("::set-output name=provider::none")
         print("::set-output name=model::")
         print("::set-output name=skip::true")
         print(f"::set-output name=reason::Model Router crashed: {e}")
+
         if "GITHUB_OUTPUT" in os.environ:
             try:
                 with open(os.environ["GITHUB_OUTPUT"], "a") as go:
-                    go.write(f"provider=none\nmodel=\nskip=true\nreason=Model Router crashed: {e}\n")
+                    go.write("provider=none\n")
+                    go.write("model=\n")
+                    go.write("skip=true\n")
+                    go.write(f"reason=Model Router crashed: {e}\n")
             except Exception:
                 pass
diff --git a/tests/test_multi_ai_ci.py b/tests/test_multi_ai_ci.py
new file mode 100644
index 000000000..814dbc0f7
--- /dev/null
+++ b/tests/test_multi_ai_ci.py
@@ -0,0 +1,108 @@
+import os
+import json
+import tempfile
+import shutil
+import pytest
+from pathlib import Path
+
+# Add multi-ai-cli directory to python path for imports
+import sys
+sys.path.insert(0, str(Path(__file__).parent.parent / "multi-ai-cli"))
+
+from ci_mode import run_ci, main
+
+def test_run_ci_permissions():
+    # Create secure temporary directory
+    temp_dir = tempfile.mkdtemp()
+    try:
+        # We will mock the DeepSeekBackend send_message to return a mock string
+        # so that it executes completely offline without network requirements
+        import backends.deepseek as ds_mod
+        original_send = ds_mod.DeepSeekBackend.send_message
+        ds_mod.DeepSeekBackend.send_message = lambda self, msg, context: "Mocked Code Review: LGTM!"
+
+        # Mock os.environ to contain GITHUB_EVENT_PATH
+        event_path = Path(temp_dir) / 'event.json'
+        event_data = {
+            'action': 'synchronize',
+            'pull_request': {'number': 137},
+            'repository': {'full_name': 'test/repo'}
+        }
+        with open(event_path, 'w') as f:
+            json.dump(event_data, f)
+
+        os.environ['GITHUB_EVENT_PATH'] = str(event_path)
+        os.environ['OPERATOR_TOKEN'] = 'test_token'
+
+        # Call run_ci
+        result = run_ci(
+            event=event_data,
+            workspace=temp_dir,
+            operator_token='test_token'
+        )
+
+        assert result is not None
+        assert result['provider_used'] == 'deepseek'
+        assert result['event'] == 'synchronize'
+        assert len(result['actions']) == 1
+        assert "Mocked Code Review" in result['actions'][0]['summary']
+
+        # Restore original function
+        ds_mod.DeepSeekBackend.send_message = original_send
+
+    finally:
+        shutil.rmtree(temp_dir)
+
+def test_ci_output_permissions(monkeypatch):
+    # Mock ci_mode main execution
+    temp_dir = tempfile.mkdtemp()
+    try:
+        cache_dir = Path(temp_dir) / 'cache'
+        event_path = Path(temp_dir) / 'event.json'
+        event_data = {
+            'action': 'opened',
+            'pull_request': {'number': 137},
+            'repository': {'full_name': 'test/repo'}
+        }
+        with open(event_path, 'w') as f:
+            json.dump(event_data, f)
+
+        monkeypatch.setenv('GITHUB_EVENT_PATH', str(event_path))
+        monkeypatch.setenv('OPERATOR_TOKEN', 'test_token')
+
+        import backends.deepseek as ds_mod
+        monkeypatch.setattr(
+            ds_mod.DeepSeekBackend,
+            "send_message",
+            lambda self, msg, ctx: "Mocked output!"
+        )
+
+        # Execute main with argparse arguments
+        monkeypatch.setattr(sys, "argv", [
+            "ci_mode.py",
+            "--workspace", temp_dir,
+            "--cache-dir", str(cache_dir)
+        ])
+
+        # Run main()
+        # Since deepseek_output.json is written in the current working directory,
+        # we change directory temporarily, or mock its output path if needed.
+        # Let's mock write location to prevent side-effects on current repo
+        original_cwd = os.getcwd()
+        os.chdir(temp_dir)
+        try:
+            main()
+
+            # Verify cache_dir exists with 0o700
+            assert cache_dir.exists()
+            assert (os.stat(cache_dir).st_mode & 0o777) == 0o700
+
+            # Verify deepseek_output.json exists with 0o600
+            output_file = Path('deepseek_output.json')
+            assert output_file.exists()
+            assert (os.stat(output_file).st_mode & 0o777) == 0o600
+        finally:
+            os.chdir(original_cwd)
+
+    finally:
+        shutil.rmtree(temp_dir)