一键在 Manus 中运行任何 Skill

debug-negotiation-summary-silent

星标121

分支20

更新时间2026年6月18日 07:33

Diagnose why the Edge/AgentVillage afternoon negotiation-summary cron "triggered but sent no Telegram message" on a tenant sidecar. Use when a user manually triggered the negotiation summary and nothing arrived. Walks the dedup-by-design check (heartbeat-state reportedCompletedIds), a non-mutating sidecar probe of the data script, and rules out OpenRouter/Telegram/index-MCP causes. The usual cause is correct anti-spam dedup, not a bug.

安装

用 Codex 或 Claude 帮你安装复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。

在 Manus 中运行

来源

indexnetwork

indexnetwork/index

打开 GitHub 仓库查看创作者相关仓库

下载

在 Manus 中运行

debug-negotiation-summary-silent

The negotiation-summary cron (Edge — negotiation summary, ~14:00 staggered, delivered via --deliver telegram) is silent by design when there's nothing new to report. A user who manually re-triggers it after a previous successful run will usually get no message — and that is correct, not broken.

The script is skills/index-network/scripts/summarize-negotiations.ts (in the Edge-City/agentvillage submodule). It buckets negotiations into needsAttention (active + your turn), waiting (active + their turn), newlyResolved (completed, updated ≤7d, not already reported), and emits [SILENT] when all three are empty. Completed negotiations already announced are tracked in memory/heartbeat-state.json → negotiationSummary.reportedCompletedIds, so the same concluded thread is never re-sent.

First: rule out the scary causes (control-plane, read-only)

Get control-plane creds from its Railway vars (CONTROL_PLANE_API_KEY, domain control-plane-production-b752.up.railway.app) — see railway-mcp-edge-city for the snake_case IDs. Then for the tenant (look up by email):

CP="https://control-plane-production-b752.up.railway.app"; KEY="<CONTROL_PLANE_API_KEY>"
curl -sf "$CP/tenants" -H "Authorization: Bearer $KEY" | jq -r '(.tenants//.results//.)[]|[.id,.email]|@tsv' | grep <email>
curl -sf "$CP/tenants/<tenantId>?detail=1" -H "Authorization: Bearer $KEY" | jq '{status,sidecarReady,gatewayReady,telegramUserId,approvedTelegramUsers,openrouter}'

Healthy tenant = status:"live", sidecarReady/gatewayReady:true, openrouter.disabled:false with credit, Telegram user id present in approvedTelegramUsers. If those are green, it is not a deploy/credits/Telegram-config failure.

Railway deploy logs for these sidecars are unreliable for this: the stream only shows gateway/process noise and is often frozen hours behind; the cron session stdout is not surfaced. Don't try to confirm the run from railway_get_logs — probe the script directly instead.

Then: probe + reset on the sidecar (decisive)

The Railway MCP has no exec/file tool, and pool sidecars (hermes-pool-<hex>) have only an internal domain. So the user (or anyone with the Railway CLI) must shell in. Find the tenant's sidecar via railwayServiceId in the ?detail=1 payload.

railway ssh        # agentvillage-controlplane → production → hermes-pool-<hex>
cd /opt/data

# 1. What does the cron actually see? Throwaway state = no dedup, no mutation:
bun skills/index-network/scripts/summarize-negotiations.ts --state-file /tmp/probe.json; echo
#   [SILENT]  -> nothing to report (dedup, or index-MCP fetch failed) — data side
#   {JSON...} -> data is fine; problem is delivery or the trigger ran a DIFFERENT cron

# 2. Confirm the real dedup state:
jq '.negotiationSummary.reportedCompletedIds' memory/heartbeat-state.json

# 3. To force a re-render of an already-reported (connected) thread, clear ONLY that list:
jq '.negotiationSummary.reportedCompletedIds = []' memory/heartbeat-state.json > /tmp/h.json \
  && mv /tmp/h.json memory/heartbeat-state.json

Then re-trigger and the summary lands (verified: clearing the list re-qualifies a connected thread as newlyResolved). Emptying the array (vs deleting the file) preserves sibling state like the digest dedup.

Gotchas

Wrong cron: the tenant dashboard "Generate" button triggers the morning digest, NOT the negotiation summary. Confirm how the user triggered — if it's the dashboard, that alone explains the missing negotiation summary.
Timing: the cron composes via an LLM session then delivers; allow 1–2 min before declaring it silent.
[SILENT] is success, not an error — both an empty bucket set and an index-MCP fetch failure produce it (the script swallows fetch errors → silent: "mcp-fetch-failed").
Don't run the script against the real memory/heartbeat-state.json for a probe — it persists reportedCompletedIds and re-arms the dedup. Use a /tmp state file.

See also: railway-mcp-edge-city (sidecar IDs + the snake_case gotcha), audit-digests (digest card quality).

同仓库更多 Skills

同仓库

finish-pr

indexnetwork/index

Finish a pull request end-to-end from the canonical/root session: investigate PR health (base freshness, checks, review threads, version bumps, local builds), hand off any needed worktree changes to the worktree session as a fix prompt — never mutating the worktree from this session — merge the PR after explicit confirmation, verify post-merge GitHub/Railway deployment health, and close or update related GitHub and Linear issues. Use when the user says a PR is ready to finish, ship, merge, or close out.

2026-07-17121

run-protocol-evals

indexnetwork/index

Run the packages/protocol eval harnesses correctly — the live LLM harnesses (bun run eval:matching/hyde/premise/profile/opportunity/clarification, which need OPENROUTER_API_KEY auto-loaded from root .env.test) versus the provider-free CI gate (bun run eval:verify, which strips credentials and never calls a model). Use when asked to "run the evals", verify eval suites, add a new eval harness, update a baseline, or when eval:verify fails on an unlisted directory or a baseline-coverage spec after adding a corpus case.

2026-07-17121

bump-agentvillage-submodule-pointer

indexnetwork/index

Ship a change to the Edge-City agentvillage submodule end-to-end and bump the index monorepo's submodule pointer, without initializing the submodule in a worktree. Use after merging a PR in Edge-City/agentvillage (or when the pointer at packages/edge-city/agentvillage is stale) — covers the branch-in-submodule → PR-to-Edge-City → pointer-bump-PR-to-dev flow, the git update-index --cacheinfo 160000 plumbing trick for committing a gitlink from a worktree with an empty submodule dir, and restoring the root submodule to a clean detached checkout.

2026-07-17121

flag-rollout-consistency

indexnetwork/index

Ship and flip feature flags consistently across ALL env surfaces in the index monorepo — .env.example (committed, commented docs), root .env.development (gitignored local mirror of Railway dev), Railway dev service variables (Railway MCP), and startup.env.ts registration. Use when adding a new env-gated feature flag, enabling/disabling a flag on dev, or when a flag behaves differently locally vs on Railway. Covers the ship-dark→flip order, railway_set_variables gotchas (snake_case ids, map-shaped variables, auto-redeploy), and why the root-dev-guard warning on .env.development edits is safe to ignore.

2026-07-17121

git-worktree-workflow

indexnetwork/index

Create and operate git worktrees in the index monorepo using the project's `bun run worktree:*` helpers instead of raw git. Use whenever you need an isolated branch checkout to make changes (the canonical root must stay on dev and is read-only for the assistant, enforced by the root-dev-guard extension), or when a worktree is missing env files, node_modules, or git hooks, or when a tool/bash call is blocked for touching the canonical root, or when `git commit`/`git rebase` fails with "gpg: signing failed: Inappropriate ioctl for device" in a non-interactive shell. Covers the dashed-folder / slashed-branch naming convention, why `worktree:setup` is mandatory after creating a worktree, and how to disable GPG signing worktree-locally without touching the user's repo-wide signing config.

2026-07-17121

open-release-pr

indexnetwork/index

Open a dated release branch from dev and create a GitHub pull request into main with a generated changelog from git log and git diff. Use when preparing a dev-to-main release PR or when the user asks to cut/open a release PR.

2026-07-17121

name

debug-negotiation-summary-silent

description

debug-negotiation-summary-silent

First: rule out the scary causes (control-plane, read-only)

CP="https://control-plane-production-b752.up.railway.app"; KEY="<CONTROL_PLANE_API_KEY>"
curl -sf "$CP/tenants" -H "Authorization: Bearer $KEY" | jq -r '(.tenants//.results//.)[]|[.id,.email]|@tsv' | grep <email>
curl -sf "$CP/tenants/<tenantId>?detail=1" -H "Authorization: Bearer $KEY" | jq '{status,sidecarReady,gatewayReady,telegramUserId,approvedTelegramUsers,openrouter}'

Railway deploy logs for these sidecars are unreliable for this: the stream only shows gateway/process noise and is often frozen hours behind; the cron session stdout is not surfaced. Don't try to confirm the run from railway_get_logs — probe the script directly instead.

Then: probe + reset on the sidecar (decisive)

railway ssh        # agentvillage-controlplane → production → hermes-pool-<hex>
cd /opt/data

# 1. What does the cron actually see? Throwaway state = no dedup, no mutation:
bun skills/index-network/scripts/summarize-negotiations.ts --state-file /tmp/probe.json; echo
#   [SILENT]  -> nothing to report (dedup, or index-MCP fetch failed) — data side
#   {JSON...} -> data is fine; problem is delivery or the trigger ran a DIFFERENT cron

# 2. Confirm the real dedup state:
jq '.negotiationSummary.reportedCompletedIds' memory/heartbeat-state.json

# 3. To force a re-render of an already-reported (connected) thread, clear ONLY that list:
jq '.negotiationSummary.reportedCompletedIds = []' memory/heartbeat-state.json > /tmp/h.json \
  && mv /tmp/h.json memory/heartbeat-state.json

Gotchas

Wrong cron: the tenant dashboard "Generate" button triggers the morning digest, NOT the negotiation summary. Confirm how the user triggered — if it's the dashboard, that alone explains the missing negotiation summary.
Timing: the cron composes via an LLM session then delivers; allow 1–2 min before declaring it silent.
[SILENT] is success, not an error — both an empty bucket set and an index-MCP fetch failure produce it (the script swallows fetch errors → silent: "mcp-fetch-failed").
Don't run the script against the real memory/heartbeat-state.json for a probe — it persists reportedCompletedIds and re-arms the dedup. Use a /tmp state file.

See also: railway-mcp-edge-city (sidecar IDs + the snake_case gotcha), audit-digests (digest card quality).