channel-debug-core

Channel Debug Core

Use this skill for any Slack, Telegram, Discord, or WhatsApp delivery issue.

Non-Negotiables

Before proposing a fix, produce:

1. Deployment-state proof.
2. Runtime path diagram with each edge marked unknown, verified-good, or verified-bad.
3. Hypothesis table with fastest falsifier and status.
4. Channel Specialist Handoff.

Do not write "most likely" twice. Gather direct evidence.

Parallel First Rule

For live deployments, channel debugging starts with parallel lanes. The parent agent creates one artifact root and launches these lanes before any patch proposal:

No fix may be proposed until the merge gate is satisfied:

1. Each lane has returned a handoff, or the lane is marked unavailable with the attempted fallback.
2. At least one runtime edge is verified-bad.
3. Contradictory lane evidence has been added to the hypothesis table.
4. Prior known fixes have been compared against current evidence.

Project Targeting Preflight

Before any vercel, npx sandbox, or npx workflow investigation, prove the command is targeting the intended Vercel project and team.

Always inspect .vercel/project.json and compare it to the incident target. If they differ, do not rely on inferred project context. Use explicit project/team/deployment targeting and record the mismatch in the handoff.

Known sharp edge: this repo may be linked to release 45 in .vercel/project.json (projectName: vercel-openclaw-45, projectId: prj_ag6Uzj9b4deNK98jVS5SO0H2Shmx). For release 46, use explicit target context (projectName: vercel-openclaw-46, projectId: prj_JSzrXyJiMzT6F7BUA76Naa7nkjRa, team vercel-internal-playground) rather than trusting .vercel inference.

If a tool prints that it inferred the project from .vercel, treat the output as suspect until the inferred project is proven to match the incident target. This is especially important for npx workflow, which can silently inspect the wrong project when run from a repo linked to a different release.

Required Admin Surfaces

Collect these first:

GET /api/admin/why-not-ready
GET /api/channels/summary
GET /api/admin/sandbox-diag
GET /api/admin/logs

Save the raw JSON before continuing. Logs are a ring buffer and can evict important events.

Readiness Triad

Report these separately:

• route-ready: platform route/native route appears registered.
• native-accepted: lastForward.ok true with classification:"accepted", or equivalent handler acceptance evidence.
• user-visible-reply: real user saw a Slack/Telegram/Discord/WhatsApp reply or status update.

Never collapse these into connected.

Deployment Proof

Run or request equivalent evidence:

git rev-parse HEAD
git ls-remote origin main
curl "$URL/api/admin/sandbox-diag"

If the deployed runtime cannot be tied to the source being read, say so and stop code-level conclusions.

Evidence Artifact Rule

Write evidence under:

.agent-runs/channel-debug/<timestamp>/<channel>/

Do not commit runtime evidence. Redact admin secrets, bypass secrets, bot tokens, webhook secrets, and platform access tokens.

Release Env Files

When the incident references a release-specific env file, such as .env.45, use it only to investigate that release's live run.

Allowed uses:

• deployment URL / app URL
• Vercel project/team/deployment context
• Vercel token for logs/workflow/sandbox CLI
• ADMIN_SECRET
• Deployment Protection bypass
• platform credentials required to verify the specific channel run

Rules:

• Source the file; do not print it.
• Do not copy it into artifacts.
• Do not include raw values in handoffs.
• Record only the file name and variable names used.
• Treat env files as credential/context input, not deployment proof. Deployment proof still comes from git, project targeting proof, and live runtime signals.

Sandbox Runtime Probe

The sandbox lane must first identify the actual Vercel Sandbox runtime ID. Use /api/admin/sandbox-diag and /api/channels/summary before CLI access.

If the ID is oc-prj*, prj_*, OPENCLAW_INSTANCE_ID, or VERCEL_PROJECT_ID, do not pass it as <sandbox_id> to sandbox exec or sandbox connect.

When npx sandbox / sandbox works, collect read-only evidence:

sandbox exec "$SANDBOX_ID" sh -lc '
  set -eu
  echo "== process =="
  ps -eo pid,ppid,comm,args | grep -E "[o]penclaw|[n]ode" || true
  echo "== ports =="
  (ss -ltnp || netstat -ltnp) 2>/dev/null | grep -E ":(3000|8787)\\b" || true
  echo "== openclaw logs =="
  tail -200 /tmp/openclaw/openclaw-*.log 2>/dev/null || true
'

For config, print only shape, never secrets:

sandbox exec "$SANDBOX_ID" node - <<'NODE'
const fs = require("fs");
const p = "/home/vercel-sandbox/.openclaw/openclaw.json";
const j = JSON.parse(fs.readFileSync(p, "utf8"));
console.log(JSON.stringify({
  channels: Object.keys(j.channels || {}),
  pluginAllow: j.plugins?.allow || null,
  hasSlack: Boolean(j.channels?.slack),
  hasTelegram: Boolean(j.channels?.telegram),
  hasDiscord: Boolean(j.channels?.discord),
  hasWhatsApp: Boolean(j.channels?.whatsapp),
}, null, 2));
NODE

If CLI access fails because the ID is incompatible or auth cannot resolve the sandbox, save the CLI failure and switch to the app admin SSH/exec fallback. The fallback must run the same read-only probes.

Workflow State Probe

The workflow lane must inspect Vercel Workflow state before treating the workflow path as healthy or broken:

npx workflow inspect runs --backend vercel > "$ART/workflow/runs.txt" 2> "$ART/workflow/runs.stderr.txt"

Correlate by channel, incident time, request ID, delivery ID, platform event ID, and drainChannelWorkflow.

Collect workflow run ID, status, current/last step, retry count or repeated step evidence, terminal error, whether native forward attempts happened, and whether channels.forward_outcome was written.

If npx workflow is unavailable or cannot inspect Vercel state, save the failure and fall back to Vercel logs filtered to workflow/channel events, /api/admin/logs, and /api/channels/summary.lastForward. Do not use local workflow data as evidence for production behavior.

Prior Fix Comparison

Before proposing a patch, compare current evidence to known prior fixes and failure signatures:

A repeated operator-facing message is not enough. Match on runtime signals: status codes, route probes, process/log evidence, Workflow state, and lastForward.

Fallback When Live Access Is Unavailable

If you cannot call the deployed endpoints, produce a static code-path audit and mark runtime edges unknown. Do not claim runtime behavior from code alone.