Tiered Help from the Command Taxonomy
The Broader Goal at the End of the Previous PR
PR #2010 (2026-10-04) wired the OpenRouter dispatch gateway — expanding harness capability to include frontier models beyond the core five. The goalpost at that point was surface polish: the core orchestration layer was complete, and user-facing commands needed refinement to match the expanded scope.
What Changed: Surfacing the Taxonomy to Users
The COMMAND_TAXONOMY (previously internal-only, used for telemetry and validation) became the source of truth for human-facing command discovery. This PR took that read-only step further by building a tiered help system directly from taxonomy entries, grouped into four command tiers:
- Core (8 commands) — the essential daily set:
init,dispatch,status,jobs,decide,pr check,exec,doctor - Workflow (maturity tier ≤ 2) — intermediate use cases
- Advanced/Admin (tier 3 and latent) — operator-level utilities
- All (
synlynk help --all) — complete command catalog
No behavior change to any existing command. The existing --help output and top-level help remain
untouched. This is pure read-side surface.
Technical Implementation
The implementation is minimal — under 60 lines of new taxonomy code plus CLI wiring:
entries_for_help(group, include_all=False)filters taxonomy entries by groupformat_tiered_help(group, include_all=False)renders compact, human-readable output with trigger phrases from the taxonomy as inline hintssynlynk helpcommand added to the top-level CLI withgroupargument (optional, defaults to "core") and--allflag- Early exit in
main()to handle help before the normal parser boots, avoiding import overhead
Tests confirm all taxonomy groups render correctly and the core tier matches the canonical list.
Guardrail
Issue #1973 (sibling to #1974) captured baseline advanced-command usage metrics before this shipped. Post-launch monitoring will compare actual usage (from telemetry) against that baseline — the guardrail is that advanced-command usage must not drop after this help system ships, confirming the new surface didn't inadvertently obscure the commands power users relied on.
What Was Achieved
One surface command now serves as the entry point to the entire command space, removing the need to
consult separate documentation for command discovery. The hierarchy (core → workflow → advanced)
mirrors the maturity model already embedded in COMMAND_TAXONOMY, so it requires zero additional
maintenance — each taxonomy entry's maturity_tier automatically slots it into the right tier.
New Goalpost
The command taxonomy is now the bridge between what the system can do (execution capabilities
tracked in capability_ratings/cost_entries) and what users can discover (surface via
tiered help). The next step is expanding that bridge: using trigger phrases and taxonomy hints to
power fuzzy command search, AI-assisted command recommendation, and shell completion scaffolds —
all derived from the same taxonomy source, keeping discovery, routing, and execution consistently
aligned as the feature set grows.
No deploy or migration required. The PR is backward-compatible; all existing commands work exactly as before.