Agent-Skills.md

Agent Skills: Bitbucket Pipeline Watch — wake-on-completion, diagnosis-in-hand

Use when an agent needs to WAIT for a Bitbucket Cloud pipeline/build to finish and act on the outcome — instead of fixed-interval polling. Backgrounds a poll-until-done loop that exits the instant the build COMPLETES, so the harness re-invokes the agent on the real event; on FAILURE it returns the redacted failure diagnosis (failed steps + error-relevant log tail) already baked in. Pairs with the maos-mcp-hub atlassian_bitbucket gateway.

UncategorizedID: ekson73/multi-agent-os/bitbucket-pipeline-watch

Repository

ekson73/multi-agent-os
ekson73License: MIT

Install this agent skill to your local

pnpm dlx add-skill https://github.com/ekson73/multi-agent-os/tree/HEAD/skills/bitbucket-pipeline-watch

Skill Files

Browse the full folder contents for bitbucket-pipeline-watch.

Download Skill

Loading file tree…

skills/bitbucket-pipeline-watch/SKILL.md

Skill Metadata

Name
bitbucket-pipeline-watch
Description
Use when an agent needs to WAIT for a Bitbucket Cloud pipeline/build to finish and act on the outcome — instead of fixed-interval polling. Backgrounds a poll-until-done loop that exits the instant the build COMPLETES, so the harness re-invokes the agent on the real event; on FAILURE it returns the redacted failure diagnosis (failed steps + error-relevant log tail) already baked in. Pairs with the maos-mcp-hub atlassian_bitbucket gateway.

Bitbucket Pipeline Watch — wake-on-completion, diagnosis-in-hand

§0 — BEING > Rules

Serves the operator's intent: stop burning context-cache on ScheduleWakeup polling and stop the 3-call failure dance (get_steps → auto_diagnose → fetch log). If it obstructs delivering value NOW, skip it and proceed. ⛔ The API token is never echoed — secret-safety is non-negotiable.

The problem it solves

An agent that wants to "act when the pipeline finishes" has two bad options:

  1. Fixed-interval ScheduleWakeup — wakes every N seconds, pays a context cache-miss each time, and still has to go fetch the failure details.
  2. A true inbound webhook — impossible: a local agent has no public HTTP endpoint, and the harness has no "inbound POST → re-invoke agent" bridge.

The viable equivalent: a backgrounded poll-until-done loop. The harness re-invokes the agent when a backgrounded Bash command exits. So a script that blocks until the build ends — and prints the verdict + (on failure) the diagnosis before exiting — gives event-driven wake-on-completion with the answer already in hand. No public endpoint, no cache-miss spam, no follow-up calls.

When to use

  • You triggered/pushed a Bitbucket pipeline and need to act on green/red (merge gate, PDCA iterate, hand-off).
  • You are in a /quiesce/auto-pilot loop monitoring a CI run.
  • You want the failure logs automatically if it fails, not just a status bit.

When NOT to use: you only need a one-shot current status → call mcp__maos-mcp-hub__atlassian_bitbucket pipeline get/get_steps directly (synchronous, no wait). The build is already done → just read it.

How to use

# Background it — the harness wakes you when it exits (on COMPLETED or timeout):
bin/bb-pipeline-watch.sh --build 1363 --repo vks-jss-sales-api &   # run_in_background:true
# or watch the latest build on a branch:
bin/bb-pipeline-watch.sh --latest --branch feature/x --repo vks-jss-sales-api
# one-shot sanity check (no loop):
bin/bb-pipeline-watch.sh --build 1363 --repo vks-jss-sales-api --once

On wake, read the command's output file:

  • === BUILD <n> COMPLETED: SUCCESSFUL === → act on green.
  • === BUILD <n> COMPLETED: FAILED === followed by FAILURE DIAGNOSIS (redacted) → the failed step(s) + the error-relevant log tail are already there; decide the fix without extra calls.

Contract

  • Creds: read from the maos-mcp-hub .env (BITBUCKET_API_TOKEN, Bearer) in a subshell; override with --env-file / BB_WATCH_ENV_FILE. The token is used only in an Authorization header and never printed.
  • Redaction: all log output passes through redact() (masks ASIA…/AKIA… keys, long base64 secrets, *_token/*secret*/password values). Anti-theater: a leaked secret in a "helpful" log is worse than no log.
  • Exit: 0 on a reached verdict OR timeout (clean background completion → harness wake); non-zero only on setup error (missing token / bad args).
  • Bounds: --interval (default 45s), --max-polls (default 80 ≈ 60min). Deterministic — no model judgment in the loop.

Anti-patterns

  • ❌ Fixed-interval ScheduleWakeup polling for a build the harness could wake you on. ❌ Calling get_steps→diagnose→fetch log by hand on every failure (this bakes it in). ❌ Echoing the token or an un-redacted log. ❌ Using it for a one-shot status (use the MCP gateway directly).

Refs

  • Pairs with mcp-tools/maos-mcp-hub atlassian_bitbucket gateway (pipeline get/get_steps/auto_diagnose).
  • Pattern origin: vks-jss-sales-api PR#85 (build 1362 FAILED → needed get_steps+diagnose+log by hand; this skill collapses that into one backgrounded wait). Cross-link [[bitbucket-pipeline-watch]].
  • Mechanism: Claude Code re-invokes the agent on backgrounded-Bash exit (the "wake me when it finishes" equivalent of a webhook).