live-bugbash
Skill5 starsChanged 17 days agoApache-2.0View on GitHub
Safe-by-default Symphony for Trello bug bash. Uses fake Trello, fake Codex, and fake/local GitHub unless explicitly asked for a real live bug bash; the phrase "hardened host" enables run-scoped danger-full-access and Codex dangerous-bypass coverage. Covers SPEC.md conformance, installers, CLI options, workflow parameters, parallelism, access modes, and sanitized local issue drafts. It never publishes issues to GitHub; use $publish-bugbash-issues later after human review to create new issues or comment on duplicates.
In this repository3 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- publish-bugbash-issuesSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
18 rules. About 0 are common, 11 are distinctive, and 7 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈11
Distinctive · estimate
This author's own choices. Worth a look.
7
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 84
Before using RUN_ID, RUN_ROOT, ISSUE_DRAFT_DIR, or PREVIOUS_RUN_ID in any filesystem path, validate that each run ID is one safe path …
- Very specific: line 98
Read references/integration-modes.md, references/host-profiles.md, and references/safety-and-isolation.md before doing any setup.
- Very specific: line 115
When the goal affirmatively says without fakes, no fakes, REAL_INTEGRATIONS=all, do a real live bugbash, or run against real services, set:
- Distinctive: line 146
If the goal contains conflicting integration or host-profile instructions, prefer explicit key-value settings over natural-language phrases.
- Distinctive: line 150
When this skill is used for an actual bug bash, read these files before starting mutable work:
- Very specific: line 173
Run an evidence-based bug bash against the latest TARGET_BRANCH of TARGET_REPO until one of the stop conditions is met.
- Distinctive: line 185
Never mutate existing GitHub repositories, issues, pull requests, settings, secrets, collaborators, billing, workflows, releases, or …
- Distinctive: line 186
Never read or write unrelated host data.
- Distinctive: line 187
Never intentionally wipe, stress, credential-scan, broad-delete, globally reconfigure, change shell profiles, mutate package managers, …
- Distinctive: line 197
Stop at the earliest of:
- Distinctive: line 206
Do not sleep, poll timestamps, run no-op commands, or keep a background terminal open merely to accumulate no-new-findings time.
- Very specific: line 214
Before exploring, run the repeatable harness from ../../../docs/live-bugbash.md once with the resolved modes.
- Distinctive: line 233
Use this loop throughout the run:
- Very specific: line 235
Check the current host system time and compare it to RUN_END_SYSTEM_TIME.
- Distinctive: line 238
Create only run-scoped Trello, local, fake, and GitHub sandbox artifacts required for the scenario.
- Distinctive: line 242
Search existing local issue drafts and prior-run issue drafts for duplicates.
- Distinctive: line 243
Create or update a local sanitized issue draft for each unique confirmed issue.
- Very specific: line 244
Update coverage-ledger.md, progress.md, and run-owned registries immediately.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Live bug bash skill for Symphony for Trello
Use this skill when the operator asks for a bug bash, live bug bash, integration bug hunt, SPEC.md conformance pass, installer bug bash, workflow-parameter bug hunt, fake-integration bug bash, real Trello test, real Codex app-server test, access-mode bug hunt, or hardened-host bug bash for this repository.
This skill is safe by default. A normal invocation uses fake Trello, fake Codex, and fake/local GitHub so an accidental trigger cannot create external boards, repositories, issues, pull requests, or worker sessions against real services. Real integrations and hardened-host dangerous access are used only when the goal clearly opts in.
Codex should not implicitly use this skill for ordinary implementation work. This repository includes agents/openai.yaml with implicit invocation disabled.
Minimal invocations
Safe default with fakes and conservative host access:
/goal Use $live-bugbash until <future timestamp>.Safe default with fakes but full hardened-host access-mode coverage:
/goal Use $live-bugbash until <future timestamp>. This is running on a hardened host.Real run with all integrations enabled, but without hardened-host dangerous coverage:
/goal Use $live-bugbash until <future timestamp>. Do a real live bugbash without fakes.Real run with all integrations enabled and hardened-host dangerous coverage:
/goal Use $live-bugbash until <future timestamp>. Do a real live bugbash on a hardened host.Explicit real run equivalent:
/goal Use $live-bugbash until <future timestamp>. REAL_INTEGRATIONS=all HOST_PROFILE=hardened.Mixed integration run:
/goal Use $live-bugbash until <future timestamp>. TRELLO_MODE=real CODEX_MODE=fake GITHUB_MODE=fake.Continuation run:
/goal Use $live-bugbash until <future timestamp>. PREVIOUS_RUN_ID=live-bugbash-20260629T103000Z.Advanced override:
/goal Use $live-bugbash until <future timestamp>. RUN_ID=live-bugbash-manual-2 ACTIVE_EXPLORATION_MINUTES_WITHOUT_FINDINGS=60 REAL_INTEGRATIONS=all HOST_PROFILE=hardened.Defaults
Unless the operator explicitly overrides them, use these values:
TARGET_REPO=martin-francois/symphony-trello
TARGET_BRANCH=main
RUN_ID=live-bugbash-<UTC timestamp as YYYYMMDDTHHMMSSZ>
RUN_ROOT=target/live-bugbash/<RUN_ID>
ISSUE_DRAFT_DIR=target/live-bugbash/<RUN_ID>/issues
GITHUB_SANDBOX_PREFIX=live-bugbash
ACTIVE_EXPLORATION_MINUTES_WITHOUT_FINDINGS=60
TRELLO_MODE=fake
CODEX_MODE=fake
GITHUB_MODE=fake
HOST_PROFILE=standardParse until <time> in the goal as RUN_END_SYSTEM_TIME. The end time is a work-until deadline, not just an upper bound for a quick smoke pass: keep actively bug-bashing until that time unless a stop condition below applies. Prefer timezone-qualified timestamps. If the timestamp has no timezone, treat it as the Codex host local system time and record that assumption in progress.md. If only HH:MM is supplied, treat it as today's host-local time. If the parsed end time is already in the past, stop and ask for a future timestamp rather than silently rolling it to another day.
Before using RUN_ID, RUN_ROOT, ISSUE_DRAFT_DIR, or PREVIOUS_RUN_ID in any filesystem path,
validate that each run ID is one safe path segment: ASCII letters, digits, ., _, and - only;
not empty; not . or ..; no slash, backslash, control character, shell metacharacter, or
percent-encoded separator. Resolve and normalize RUN_ROOT and ISSUE_DRAFT_DIR before creating,
reading, or deleting anything. ISSUE_DRAFT_DIR must stay under RUN_ROOT, and default run roots
must stay under target/live-bugbash/. If an explicit path override does not normalize to a
run-owned location, stop and ask for a corrected run-scoped path.
PREVIOUS_RUN_ID is optional. When present, apply the same safe-segment validation, then read only
the previous run's final report, issue drafts, and coverage ledger to avoid duplicate testing and
duplicate issue drafts. Do not inherit its mutable state.
Mode resolution
Read references/integration-modes.md, references/host-profiles.md, and references/safety-and-isolation.md before doing any setup. Resolve integration modes and the host profile before any command that could touch Trello, GitHub, Codex, workers, credentials, or host state.
The words live-bugbash or $live-bugbash by themselves do not opt into real external services or host dangerous access.
Real integrations require one of these affirmative explicit signals in the goal. Ignore the phrase
when it appears inside a denial such as do not use real Trello, without real services, or not a
real run; keep the affected mode fake when intent is ambiguous.
REAL_INTEGRATIONS=allTRELLO_MODE=real,CODEX_MODE=real, orGITHUB_MODE=real-sandboxuse real trello,use real codex, oruse real githubuse real trello, codex and githubwithout fakesno fakesdo a real live bugbashrun against real services
When the goal affirmatively says without fakes, no fakes, REAL_INTEGRATIONS=all, do a real
live bugbash, or run against real services, set:
TRELLO_MODE=real
CODEX_MODE=real
GITHUB_MODE=real-sandboxHardened-host dangerous access requires one of these affirmative explicit signals in the goal. Ignore
the phrase when it appears inside a denial such as not a hardened host, no dangerous host access,
or do not run dangerous access on host; keep HOST_PROFILE=standard when intent is ambiguous.
HOST_PROFILE=hardenedHARDENED_HOST=truehardened hosthardened systemhardened runnerthis host is hardenedhost is hardenedrun dangerous access on hostexercise dangerous access on hostrun the dangerously things on the hostsafe to use danger full access on host
When hardened-host access is enabled, set HOST_PROFILE=hardened. This means the run should include host-level dangerous-access scenarios where the resolved integration modes allow them. It includes setup-local --danger-full-access, SYMPHONY_CODEX_DANGER_FULL_ACCESS=true, workflow-authored dangerFullAccess, direct Codex --sandbox danger-full-access --ask-for-approval never, and Codex --dangerously-bypass-approvals-and-sandbox or --yolo when supported by the installed Codex and when CODEX_MODE=real for direct Codex execution.
A hardened-host signal by itself does not imply real Trello, real Codex, or real GitHub. The phrase do a real live bugbash on a hardened host opts into both real integrations and hardened-host dangerous-access coverage.
GITHUB_MODE=real-sandbox still means GitHub writes are restricted to newly created private sandbox repositories whose names include RUN_ID. It never permits writes to existing repositories, existing issues, existing pull requests, settings, secrets, organization configuration, billing, collaborators, or releases.
If the goal contains conflicting integration or host-profile instructions, prefer explicit key-value settings over natural-language phrases. If ambiguity remains, choose the safer mode for the ambiguous service or access profile and record the decision.
Required references
When this skill is used for an actual bug bash, read these files before starting mutable work:
../../../AGENTS.md../../../README.md../../../SPEC.md../../../.github/ISSUE_TEMPLATE/bug_report.yml../../../CONTRIBUTING.mdreferences/integration-modes.mdreferences/host-profiles.mdreferences/safety-and-isolation.mdreferences/runbook.mdreferences/coverage-matrix.mdreferences/issue-drafts-and-reporting.md
Also read these repository docs when present: ../../../docs/live-e2e.md,
../../../docs/operations.md, the "Allowed Host Paths And Sandbox" section in ../../../README.md,
../../../docs/agents/deployment-and-live-verification.md,
../../../docs/agents/default-workflow.md, and relevant ADRs found during the run. If a listed file
does not exist, record that as a repository-shape observation and continue with the available
equivalents.
Goal
Run an evidence-based bug bash against the latest TARGET_BRANCH of TARGET_REPO until one of the stop conditions is met.
In fake mode, use deterministic fake or local services to exercise the same product surfaces without touching external Trello, Codex, or GitHub services. In real mode, use real Trello and real Codex through codex app-server; use real GitHub only through newly created private sandbox repositories.
When HOST_PROFILE=hardened, exercise the full host access-mode matrix, including path-scoped writable roots, all-path setup flows, danger-full-access, setup-local --danger-full-access, SYMPHONY_CODEX_DANGER_FULL_ACCESS=true, workflow-authored dangerFullAccess, Codex --sandbox danger-full-access --ask-for-approval never, and Codex --dangerously-bypass-approvals-and-sandbox or --yolo when real Codex is enabled and the installed CLI exposes those flags. These tests must remain run-scoped and must not intentionally damage or globally reconfigure the host.
Find confirmed, reproducible product bugs. For each confirmed bug, create a sanitized local issue draft under ISSUE_DRAFT_DIR. Do not create, edit, label, or comment on issues in martin-francois/symphony-trello or any other existing GitHub repository. Publishing reviewed issue drafts and commenting on duplicate issues is a separate explicit workflow handled by $publish-bugbash-issues, not by this skill.
Safety model
The safety model depends on the integration modes and host profile, but these rules always apply:
- Never mutate existing GitHub repositories, issues, pull requests, settings, secrets, collaborators, billing, workflows, releases, or organization settings. This skill creates local issue drafts only.
- Never read or write unrelated host data.
- Never intentionally wipe, stress, credential-scan, broad-delete, globally reconfigure, change shell profiles, mutate package managers, change service managers, kill unrelated processes, chmod or chown unrelated paths, or inspect unrelated private data.
- Installer, uninstaller, onboarding, cleanup, repair, and destructive local-data tests must set
HOME,XDG_*,SYMPHONY_HOME, app, config, state, cache, workspace, prefix, bin, manifest, and env-file paths to run-owned locations. In real-integration runs, pass live auth through explicit auth-only credentials or minimal cleanup-required run-scoped auth copies; do not use the operator's normal homes to make live Trello/GitHub/Codex execution work. - Real Trello, when enabled, may create and archive only run-scoped disposable boards and cards.
- Real GitHub, when enabled, may write only inside newly created private sandbox repositories whose names include
RUN_ID. - Host
danger-full-accessand Codex--dangerously-bypass-approvals-and-sandboxor--yoloare allowed only whenHOST_PROFILE=hardened. This is permission to exercise product access modes, not permission to damage the machine.
Active exploration stop rule
The no-new-findings period is measured as active exploration time, not passive wall-clock waiting.
Stop at the earliest of:
RUN_END_SYSTEM_TIME- A non-recoverable global safety blocker after all safer alternatives are exhausted
- Full meaningful matrix coverage plus
ACTIVE_EXPLORATION_MINUTES_WITHOUT_FINDINGSminutes of active exploration without finding a new unique confirmed issue - Full meaningful matrix coverage plus documented exhaustion of useful follow-up work, evidence collection, cleanup verification, and reporting, but only after a deep follow-up pass; do not use this after a shallow first pass or when there are still meaningful edge cases to explore.
Active exploration means useful bug-bash work, such as testing uncovered behavior, varying workflow parameters, reproducing or falsifying suspected issues, inspecting relevant source code, collecting evidence, deduplicating, writing issue drafts, cleaning run-scoped artifacts, or writing the final report.
Do not sleep, poll timestamps, run no-op commands, or keep a background terminal open merely to accumulate no-new-findings time.
Short waits are allowed only when they are tied to concrete product behavior under test, such as Symphony poll intervals, fake or real Trello API consistency, fake or real Codex app-server startup, worker shutdown, retry backoff, rate-limit recovery, or log flushing. Record what condition is being waited for and continue as soon as the condition has been checked.
Do not stop early just because the first planned matrix pass is complete. If the main matrix is covered before RUN_END_SYSTEM_TIME, keep doing active exploration: vary inputs, run edge cases, test failure paths, inspect source boundaries, exercise live installed paths, deduplicate possible issues, improve evidence, or clean and report. The normal early-success exit is the no-new-findings rule above, measured in active exploration time.
Repeatable harness baseline
Before exploring, run the repeatable harness from ../../../docs/live-bugbash.md once with the
resolved modes. It replays the scenario manifest in ../../../scripts/live-bugbash/manifest.yml
and writes its own run root, so use a separate run id:
scripts/live-bugbash/run.sh --profile full --run-id <RUN_ID>-harness --time-budget <minutes until RUN_END_SYSTEM_TIME>m \
[--trello real] [--codex real] [--github real-sandbox] [--host-profile hardened]Pass --trello real only when TRELLO_MODE=real, --codex real only when CODEX_MODE=real,
--github real-sandbox only when GITHUB_MODE=real-sandbox, and --host-profile hardened only
when HOST_PROFILE=hardened. Then copy the harness results into this run's coverage ledger: rows
the harness reports covered count as covered, failed rows are suspected bugs to reproduce before
drafting, and skipped or not-yet-automated rows stay open for exploration. Link the harness
final-report.md from this run's final report. When exploration finds a repeatable scenario the
manifest lacks, record it in the final report as a candidate manifest row.
Working loop
Use this loop throughout the run:
- Check the current host system time and compare it to
RUN_END_SYSTEM_TIME. - Pick the next weakly covered row from the coverage ledger, not a random path.
- Design a small but meaningful scenario for that row using the resolved integration modes and host profile.
- Create only run-scoped Trello, local, fake, and GitHub sandbox artifacts required for the scenario.
- Execute through the product surface where practical: CLI, installer, packaged app, service API, workflow config, Trello adapter, Codex app-server adapter, and setup-local lifecycle commands.
- Observe CLI output, logs, HTTP responses, fake or real app-server behavior, fake or real Trello state, process state, workspace files, generated workflow files, and source code.
- Reproduce suspected bugs at least twice unless doing so would be destructive, rate-limited, or clearly unsafe.
- Search existing local issue drafts and prior-run issue drafts for duplicates. In real GitHub mode, upstream issue searches may be read-only. Do not write to upstream issues.
- Create or update a local sanitized issue draft for each unique confirmed issue.
- Update
coverage-ledger.md,progress.md, and run-owned registries immediately. - Clean up run-scoped mutable artifacts when safe, or record why they are intentionally left behind.
- Continue until a stop condition is met.
Final output
An actual run must leave at least these artifacts under RUN_ROOT:
progress.mdcoverage-ledger.mdfinal-report.mdcleanup-summary.mdor an equivalent cleanup section in the final reportissues/with sanitized local issue draftscreated-trello-boards.jsonlcreated-trello-cards.jsonlcreated-github-sandbox-repos.jsonlstarted-workers.jsonlowned-local-paths.txt- snapshots and private evidence as needed under
snapshots/andprivate-evidence/
The final report must include the stop reason, integration modes, host profile, real-service opt-in phrase, hardened-host opt-in phrase when any, target commit, coverage summary, issue draft list, blockers, cleanup status, artifacts intentionally left behind, private-context scan result, and recommended next runs.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

