workers agents: trigger argument unvalidated (case typo silently downgrades pipeline, exit 0); confidence pinned at 60% via dead profile branch

Summary

workers agents <trigger> has a genuinely good trigger→agent table — 12 triggers, all differentiated and domain-appropriate, with --json matching the text exactly. Two defects sit on top of it: the confidence/reasoning are structurally pinned, and the trigger argument is not validated at all, so a case typo silently downgrades the pipeline.

Measured on 2.1.2. (Absent on @alpha — see #208 §3.)

---

What works (so the fix doesn't disturb it)

trigger	primary	fallback	best
ultralearn	researcher, coder	planner	researcher
optimize	performance-analyzer, coder	researcher	performance-analyzer
consolidate	researcher	—	researcher
predict	performance-analyzer	researcher	performance-analyzer
audit	security-analyst, tester	reviewer	security-analyst
map	researcher	coder	researcher
preload	researcher	—	researcher
deepdive	researcher, security-analyst	coder	researcher
document	documenter, researcher	coder	documenter
refactor	coder, reviewer	researcher	coder
benchmark	performance-analyzer	coder, tester	performance-analyzer
testgaps	tester	coder	tester

These are sensible and clearly hand-curated. No [object Object], no text/JSON divergence.

---

1. Confidence is pinned at 60% and the reasoning is always "No performance history"

Every one of the 12 triggers returns:

"selection": {
  "agent": "researcher",
  "confidence": 0.6,
  "reasoning": "No performance history - using default recommendation"
}

This is a downstream consequence of #210, not an independent bug. selectBestAgent (dist/workers/worker-agent-integration.js:271-274):

const profile = agentPerformanceProfiles.get(bestAgent);
const confidence = profile
    ? Math.min(0.95, 0.5 + (profile.executionCount * 0.01) + (profile.qualityScore * 0.4))
    : 0.6;

agentPerformanceProfiles can never be populated — its only writer, recordAgentPerformance, has zero call sites in dist/ (#210). So the profile branch is dead code, confidence is always 0.6, and "Best Selection" is always just the first entry of primary.

The documented output — Confidence: 94% with Reason: Best match for learning tasks based on historical success — is not reachable by any code path in the package. Fixing #210 fixes this too; until then the "Best Selection" block reads as a ranking decision when it is a list index.

2. The trigger argument is not validated

All of these return the generic researcher recommendation at exit 0:

$ agentic-flow workers agents TOTALLY-INVENTED-TRIGGER   # exit 0
$ agentic-flow workers agents ""                         # exit 0
$ agentic-flow workers agents ULTRALEARN                 # exit 0
$ agentic-flow workers agents "ultralearn "              # exit 0
$ agentic-flow workers agents ../etc/passwd              # exit 0

The consequence is a silent downgrade, not just a cosmetic one:

ultralearn  → file-discovery → pattern-discovery → vectorization → summarization
ULTRALEARN  → file-discovery → summarization
TOTALLY-INVENTED-TRIGGER → file-discovery → summarization

A case typo produces output byte-identical in shape to a real trigger, while quietly losing the 4-phase pipeline — and there is no signal distinguishing "unknown trigger" from "this trigger has a minimal pipeline". workers triggers already enumerates the 12 valid values, so validating against that list is cheap.

This is inconsistent with sibling commands in the same group, which handle unknown arguments correctly:

• workers native nonsense-type → Unknown native worker type + exit 1
• workers status worker-INVENTED-999 → Worker not found + exit 1
• workers benchmark --type nonsense → Unknown benchmark type + exit 1

Minor, same cause: the unrecognised string is echoed verbatim into the memory pattern (../etc/passwd/{topic}/{phase}). These are namespace strings rather than filesystem paths, so this is an unvalidated echo rather than a traversal — but it is another place the bad input passes through unchecked.

3. The documented example matches on no field

field	documented	actual
Primary	researcher, coder, analyst	researcher, coder
Fallback	reviewer, architect	planner
Pipeline	discovery → analysis → pattern-extraction → storage	file-discovery → pattern-discovery → vectorization → summarization
Memory	{trigger}/{topic}/{timestamp}	ultralearn/{topic}/{phase}
Confidence	94%	60%
Reason	Best match … based on historical success	No performance history - using default recommendation

There is no analyst or architect agent in the capability table, and the memory pattern's last segment is {phase}, not {timestamp}.

Environment

agentic-flow 2.1.2 (global), Node 24, macOS 15.6, run in an empty scratch directory.

Start at dist/workers/worker-agent-integration.js:271-274 and inspect the workers agents argument handling, comparing it with workers native, status, and benchmark validation. Reproduce the listed invalid-trigger commands and verify that valid trigger output remains unchanged while invalid, empty, case-mismatched, and path-like values produce an error and exit 1; review #210 separately for the confidence behavior.