trigger-my-training

A ground-first reflex for coding agents. On a complex or irreversible request, the agent stops, reconstructs the domain's reality, treats its own training as a stale hypothesis, probes the live system, and is hard-blocked from the destructive step until it has grounded.

The trigger is the agent's own judgment, not a hand-authored keyword list — so it works in any domain. The plugin's thesis is its mechanism: trigger the model's training, don't encode static rules a human guessed at.

[!IMPORTANT] The core insight, after an adversarial review that killed the naive version: an agent's training is always stale. So the reflex does not use recall as the answer — it uses recall to generate the list of things to verify, then runtime overrides recall. That one reframing is what makes grounding help instead of hurt.

Install

# load locally from a clone (no marketplace needed)
claude --plugin-dir /path/to/trigger-my-training

Or install from the marketplace:

/plugin marketplace add 88plug/trigger-my-training
/plugin install trigger-my-training@trigger-my-training

Optional status-line badge (⏚ TMT:armed / :grounded):

bash /path/to/trigger-my-training/install.sh

What it does (60-second version)

Ask an agent to "deploy a VM to Proxmox" and it tends to barrel into qm create — encoding stale defaults and skipping the prerequisites that actually break the deploy. This plugin intercepts that pattern:

Recognize (ground-first skill, model-elected) — the agent judges, from its own understanding, whether a request is complex/irreversible enough to ground — in any domain, not a keyword list. (Measured: precision 1.0 / recall 0.97 across infra + diverse domains, vs a keyword classifier's 0.28 recall off-infra.)
Reconstruct — split what you KNOW from what you're ASSUMING, emit a Grounding Brief (decision points, silent failure modes, unknowns tagged PROBE / ASK / ASSUME), and verify your three riskiest assumptions.
Enforce (PreToolUse hook, self-arming) — a destructive action is denied until grounding is recorded with tmt-ground commit. No detector arms it; read-only probes and ordinary file edits are never blocked.

[!NOTE] The trigger is the model's judgment (no hand-authored keyword list — that would be the very static domain knowledge this plugin exists to replace). The gate stays deterministic, because a safety floor must not depend on the model it is gating.

Component	Surface	Role
`ground-first`	skill (+ 6 reference packs)	the soft trigger — model-elected from a keyword-free policy description; holds the Grounding Brief procedure
`tmt_enforce.py`	PreToolUse hook	hard deny on the irreversible step
`tmt_reconcile.py`	PostToolUseFailure hook	on a tool failure, reconcile the falsified assumption (predict-then-check)
`tmt_log.py` / `tmt_session.py`	PostToolUse / SessionStart	record probes / prune stale state
`tmt-ground`	bin CLI	release the gate after probing
`grounding-investigator`	agent	isolated live-probing pass for CRITICAL tasks
`tmt_statusline.sh`	status line	`⏚ TMT:armed` / `⏚ TMT:grounded` badge
`/status` `/ground` `/reset` `/brief` `/explain` `/doctor`	commands	inspect / force / disarm / brief / explain / health-check

The advisory/hard split is deliberate: the nudge raises the odds the agent grounds; the PreToolUse deny is the lever that actually holds.

Configure (at enable time)

option	default	effect
`gate_mode`	`full`	`full` (brief + gate) · `gate` (gate only) · `brief` (nudge only) · `stale` · `off`
`hard_gate`	`true`	`false` makes the gate advisory-only (keeps the nudge, drops the block)

The status line ships in bin/ but plugins can't register a main statusLine, so add it to ~/.claude/settings.json yourself — see docs/architecture.md.

Does it work? (the science)

Built falsification-first. See EXPERIMENTS.md for the full ledger; headline: