v0.13.0 — Discoverability & Accounting
Date: 2026-07-22 Tag: v0.13.0 PRs: #274–#440 (~60 PRs since v0.12.0)
One sentence: synlynk learns to explain itself — every command now has a discoverable taxonomy entry with a maturity-tiered reveal, a live selftest exercises all 59 commands end-to-end, cost accounting gets payment-model awareness and a task-boundary fence around actual spend, and dispatch routing gets capability-aware enough to stop sending GitHub-write work to agents that structurally can't do it.
What v0.13.0 Is
v0.12.0 closed the H0 measurement gate — every cost figure synlynk displays is now sourced or labeled. Seven days and roughly sixty PRs later, v0.13.0 answers the next question the platform needed to answer: does synlynk know what it can do, and does it tell you honestly? A full command taxonomy now backs FTUE, docs, and trigger phrases from one source of truth; a live selftest actually runs all 59 commands instead of trusting unit-test coverage; cost accounting learns the difference between a subscription seat and a metered API call; and dispatch routing gained a first real capability gate — can_gh_write — instead of relying on an SOP a dispatched agent has no way to enforce on itself.
This is a wide release, like v0.12.0 before it. [Unreleased] accumulated a full feature cluster without an intermediate cut — this release rolls it all up rather than retroactively splitting seven days of history.
Six interlocking epics shipped, plus two live-issue fixes:
- Command Taxonomy, Maturity-Tiered Reveal, and Trigger Registry (#303) — one registry drives FTUE, docs generation, and trigger phrases.
- Task-Boundary Cost Fence (story-615bc8f4) — shared, allowlist-gated cost display at dispatch-start and job-completion.
- Live Command Selftest (#328) — a taxonomy-driven smoke test exercising all 59 commands against a real scratch workspace.
- Capability Sweep + Industry Taxonomy (#367) — NAICS/APQC/SFIA-coded calibration scoring for agent capability.
- Payment-Model-Aware Cost Accounting (#374) — real dollars vs. pay-as-you-go-equivalent estimates, per agent.
can_gh_writeCapability Routing (#423/#426/#438) — structural enforcement of who can complete GitHub-write dispatch work.- Two live-issue fixes (#421) —
synlynk watch'sCYCLES.index("work")crash andsynlynk viz --serve's daemon-thread hang, both root-caused and shipped in this release.
The Milestones
Command Taxonomy, Maturity-Tiered Reveal, and Trigger Registry (#303)
New synlynk/taxonomy.py — COMMAND_TAXONOMY classifies all 59 CLI commands/subcommands by governs_stage, maturity_tier (FTUE / Goal / Execute / Team-Enterprise / latent), prominence, orientation_gateway, audience, trigger_phrases, and hook_event. This single registry now drives three previously-separate, previously-drifting surfaces: the FTUE wizard and synlynk launch task picker (#319), the auto-generated command reference docs (#316), and tier-scoped trigger phrases in the synlynk:start/synlynk:end session fence (#320). A pre-commit hook installed at synlynk init (#321) now gates instructions drift between the taxonomy and the generated CLAUDE.md/GEMINI.md/AGENTS.md content, closing the class of bug where the docs quietly stopped matching the CLI.
Task-Boundary Cost Fence (story-615bc8f4)
Cost/token data now surfaces at two fixed points — dispatch-start (an estimate) and job-completion (the actual) — in one shared, allowlist-gated format (synlynk/fencing.py: FenceData, render_task_fence(), is_fenced_command()), instead of being computed ad hoc at each call site. dispatch_agent() attaches an estimate to the job dict so both the direct dispatch CLI path and schedule --execute print it without duplicated estimation logic, and exec_command()'s and _format_job_summary()'s cost prints route through the same renderer. release is explicitly out of scope.
Live Command Selftest (#328)
The GTM checklist's item 1 — "deep review of every synlynk command and its testing in a live repo scenario" — got a real answer instead of a manual audit. New synlynk/selftest.py runs a taxonomy-driven smoke test against a real scratch git repo, grouped into lifecycle/free/paid scenario tiers (paid scenarios — dispatch/exec/schedule — gated behind explicit opt-in since they spend real tokens). A follow-up fix (#337) closed a real isolation gap: the paid scenarios chdir'd into the scratch workspace but never rebound synlynk.DB_PATH, so a live paid run could still write real rows into the host's actual central DB.
Capability Sweep + Industry Taxonomy (#367)
Static NAICS/APQC/SFIA lookup tables plus a migration crosswalk for legacy free-text discipline/org_domain/industry values (unmatched values tagged legacy_unmapped, never silently dropped). synlynk capability sweep runs a taxonomy-driven calibration sweep — CLI model discovery, SFIA-scoped calibration tasks, independent cross-agent verification scoring — under a configurable $10 cost guardrail. A capability_baseline.json seeds cold-start capability rows so new agents don't start from zero. synlynk pr check also gained a geometric review-cycle multiplier at merge time, rewarding clean first-pass approvals and decaying quality credit across requested-changes rounds.
Payment-Model-Aware Cost Accounting (#374)
Closes a real accuracy gap the Measurement Ledger epic didn't cover: two agents on different payment models (subscription seat vs. metered API) were being reported at the same blended token rate. payment_models config, a credit_grants ledger table, and resolve_payment_value() now branch correctly across subscription overage billing, oldest-first multi-grant consumption, and plain pay-as-you-go — with costs.md splitting rows into api-equivalent and actual-dollars-spent columns, and check_budgets()'s Budget Pulse gaining a per-agent payment-model rollup.
can_gh_write Capability Routing (#423/#426/#438)
The root problem: dispatched agents share a single GitHub identity, so gh pr review --approve always fails as self-approval, and only Claude/Grok can structurally complete GitHub-write work headless — a fact that lived only in SOP text, unenforced by code. This epic makes it structural: can_gh_write: bool on AGENT_CAPABILITY_BASELINES, a new --requires-gh-write dispatch flag, and enforcement in dispatch_agent(). A non-blocking gap remains open: reroute-target selection currently favors claude over grok due to dict insertion order, despite the SOP's "Grok only" intent — tracked, not yet fixed. The deeper identity half of #423 (distinct per-agent GitHub identities so a real gh pr review --approve becomes possible, not just a comment-review workaround) stays out of scope for this release.
Two live-issue fixes (#421)
Both reported directly during real usage, both fully root-caused before any fix shipped:
synlynk watchcrash —ValueError: 'work' is not in list. Thev0.12.0tag shippedcmd_watch()still looking up a"work"cycle name that had already been renamed out ofhud.py'sCYCLESlist under the GOVERNS seven-stage vocabulary rollout. The fix (CYCLES.index("execute")) had actually merged tomainvia #301 two days after thev0.12.0tag — it just never made it into a release until now.synlynk viz --serveimmediate exit — reported as a suspected sandbox/daemonization artifact, but reproduced identically outside the sandbox:_start_server()spawned the HTTP server on adaemon=Truethread and returned immediately, so the process (and its daemon thread) died right after printing "Serving at...". New_serve_until_stopped()blocks the main thread untilKeyboardInterrupt, then shuts the server down cleanly. Verified with a new regression test using a fake HTTP server (the sandbox blocks real localhost binds in CI).
Test Count
| Milestone | Tests |
|---|---|
| v0.12.0 (tag) | 1140 passed, 2 skipped |
| v0.13.0 | 1319 passed, 2 skipped |
What's Next
Grok's non-blocking observation on can_gh_write reroute-target ordering (favoring claude over grok by dict insertion order) is still open. Issue #423's GitHub-identity half (per-agent PAT/App so a real gh pr review --approve is possible) remains unscoped. Adding watch/viz real-launch coverage to the Live Command Selftest, flagged as a natural companion fix during this release's investigation, hasn't been scoped yet either. The security-hardening cluster (#348–#355) and the job-status/reconciliation cluster (#329–#331, #419, parked #202) remain open and untouched by this release.