team-diagnose
Overview
Evidence-first diagnosis of broken adlc-team-skills wiring. The chain has exactly five links, each checkable with one command — a diagnosis without command output is a guess, and guesses get pasted into bug reports.
Boundary: this repo owns .events.json and the boot scripts' output.
The injection side (dispatchers, generated per-agent plugins) is owned by
adlc-cli — the
docs/event-hook-contract.md file in the adlc-team-skills repo draws the
line. Diagnose up to it, then route.
When to Use
- Team context missing at session start (no
Team Context & Decisionssection). - A rule that should have matched didn't load.
- Install behaves unexpectedly after an update.
When NOT to use: for application bugs unrelated to the team-context chain, or for skill behavior debates (that's a goldset/eval question).
Core Process
Run the checks in order — stop at the first failure, fix, re-run:
| # | Symptom area | Check (run it) | Green means |
|---|---|---|---|
| 1 | Not configured | cat .adlc/init-options.json | A team_ai_directives path that exists |
| 2 | Missing tool | command -v jq | jq on PATH (skills registry needs it) |
| 3 | Broken handler | bash .agents/skills/team-boot/scripts/bash/boot.sh from the project root | The directives index on stdout, wrapped in EXTREMELY_IMPORTANT |
| 4 | Stale install | pytest tests/unit/test_generated_artifacts_sync.py -q (in the skills repo) or compare skills/**/SKILL.md vs .agents/skills/ names | Mirror matches source |
| 5 | Injection side | Generated plugin present and loaded for the agent? | If 1–4 are green and the session still lacks context → adlc-cli's territory — file there with the outputs of checks 1–4 attached |
For "a rule didn't load" specifically: confirm the session began with the index (check 3's content), then check the task actually matched the rule's descriptor — on-demand pull is match-driven, not mind-reading.
Whole-loop check: scripts/acceptance-test.sh scratch-installs and
asserts the chain end-to-end (add --live for a real agent session).
Report format
For a bug report, include: the failing check number, the command, its
verbatim output, agent name/version, and how the repo was installed
(npx skills add vs adlc-cli). No scrubbed narrative — raw
output.
Red Flags
- Concluding before running a command — every check is one command.
- "Reinstall everything" before reading check 3's output.
- Blaming a skill's content when check 1–3 failed (config, not skills).
- Filing injection-side bugs in adlc-team-skills (check 5 routes them).
All of these mean: run the checks, then report.
Verification
A diagnosis is done when:
- Checks 1–4 each ran with output captured (or 5 routed with evidence)
- The failing link is named with its command output
- The report routes to the owning repo
