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:

1. 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.)
2. 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.
3. 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:

• Detector (Exp 2, deterministic): clean separation of operational vs edit-intent on a 28-task labelled corpus — precision/recall 1.0, 0 false positives (in-sample; real-world calibration is the known debt).

• Hard gate (Exp 3, unit): blocks the mutation, allows probes + local edits, releases after grounding — 7/7.

• Landmine-catch (Exp 1, A/B + ablation, powered to 12 domains): the replicated result on claude-haiku-4-5 —

  arm	catch-rate	vs baseline
  no plugin	0.181	—
  compact Pre-Mortem Brief	0.386	~2.1×
  + Staleness Axiom alone	0.156	worse (axiom alone does nothing)
  enriched (4 composed inventions)	0.258	worse than the plain brief

  The campaign falsified its own maximalist hypothesis: a one-line "your training is stale" axiom does nothing, and composing more proven cognitive scaffolds (Tetlock calibration tags + Deming predict-then-check + Deutsch hard-to-vary) regressed the gain. The active ingredient is one 3-line structural trigger — Klein's pre-mortem + Popper's "enumerate how it breaks." Adding cognitive mass crowds it out. That result survived replication across 12 domains; an exciting single-run "win" for the enriched slate did not (Twyman's law). Honest accounting of "10×": the replicated grounding number is ~2.1×; the only literal ≥10× is the poka-yoke gate taking irreversible-action interception from ~0 to ~1.0. Full ledger: EXPERIMENTS.md, invention slate: INVENTIONS.md.

bash evals/run.sh          # all three experiments
python3 evals/detector_eval.py    # just the free deterministic one

What this is NOT

Three independent refuters killed the maximalist pitch, and the design reflects it:

• It does not claim the model "already knows" the domain — training is stale; that is the whole point.
• It does not replace human-authored skills for must-be-exact execution — it grounds the understanding layer.
• The reasoning mechanism is not novel (step-back / generated-knowledge / preflight prior art) — the contribution is the packaging and the harness-enforced gate.

Layout

.claude-plugin/   plugin.json, marketplace.json
bin/              detector, enforcer, probe-log, tmt-ground state machine, lib
hooks/            hooks.json (SessionStart, PreToolUse, PostToolUse, PostToolUseFailure)
skills/ground-first/   SKILL.md + reference/{proxmox,general}.md
agents/           grounding-investigator.md
commands/         status, ground, reset
evals/            harness.py, detector_eval.py, gate_unit_test.sh, tasks.jsonl
EXPERIMENTS.md    the falsification ledger

Contributing & security

• CONTRIBUTING.md — dev/test workflow and the bin/hook architecture
• docs/architecture.md — detect→brief→gate→release flow + findings
• docs/userconfig-design.md — the tunable options
• CHANGELOG.md · SECURITY.md

bash tests/run.sh          # unit tests (26 cases)
bash evals/run.sh          # the three experiments
claude plugin validate .   # manifest check

License

FSL-1.1-ALv2 © 2026 88plug — Functional Source License; converts to Apache 2.0 two years after each release.