← Back to Blog

v0.12.0 — Measurement & Reliability

Date: 2026-07-15 Tag: v0.12.0 PRs: #122–#267 (71 PRs since v0.11.0)

One sentence: the operational backbone gets provably reliable and provably accounted for — dispatched jobs finish their own git steps, a 5th zero-cost local agent joins the fleet, story routing gets real capability+quota+cost scoring with a fleet batch scheduler, and every dollar synlynk reports is now either structurally sourced or visibly flagged as an estimate.

What v0.12.0 Is

v0.11.0 made the agent fleet observable and permissioned. Ten days and 71 PRs later, v0.12.0 makes it trustworthy: dispatched work reliably lands in git without babysitting, story routing is a real scoring engine instead of a first-match heuristic, a zero-cost on-device agent joins the roster, and — closing out the arc that prompted this release — every cost figure synlynk shows you is either measured or clearly labeled as a guess.

This is a wider release than most. Several independently-shippable epics accumulated in [Unreleased] between v0.11.0 and today without an intermediate cut; rather than retroactively split ten days of history into several small releases, this one rolls them up together.

Five interlocking epics shipped:

Job Lifecycle Ground-Truth Verification (#126–#129) — dispatch worktree isolation, git-state-verified reconciliation, real files_touched, fail-loud migration.
Dispatch git-finalization reliability chain (#182–#199) — synlynk finishes agents' git steps for them, because they don't reliably do it themselves.
Capability Matrix Hardening + Fleet Scheduler (epic #137) — 3-stage routing (capability → quota gate → cost tie-break) and a batch scheduler that clears backlogs unattended.
Local Agent — a 5th dispatch agent running Aider over on-device oMLX inference, $0 per token.
Measurement Ledger Hardening (epic #210) — the H0 gate from Fable's strategic review: every number is structurally sourced or visibly flagged.

The Milestones

Job Lifecycle Ground-Truth Verification (#126–#129)

Four sequenced fixes closing the gap between "a dispatched job says it's done" and "a dispatched job's work is verifiably real": per-job git worktree isolation so concurrent dispatches can't collide, git-state cross-checking when a job's exit sentinel is missing (failed_unverified instead of a blind hard failure), real files_touched via git diff, and cmd_migrate() failing loud instead of printing a green banner over a 0-row import.

Dispatch git-finalization reliability chain (#182–#199)

The core finding, confirmed repeatedly across one long session: dispatched agents (Codex, Agy) frequently do real, correct, fully-tested work but don't reliably finish the meta-step — commit, push, open PR — even with explicit instructions. synlynk now does it for them: _finalize_completed_worktree_job() fires once on a job's running→terminal transition, idempotent by construction, staging everything except a hard-exclusion list, committing, pushing, and opening a PR via gh pr create if none exists. Wrapped in try/except everywhere so a git/gh hiccup never crashes reconciliation. Sibling fixes: borrowed-worktree completions no longer misattributed as zero-work, HARNESS_INTERNAL_TIMEOUT jobs auto-retry, and a job that raced its own disk writes gets one re-inspection instead of staying stuck in a stale failed forever.

Capability Matrix Hardening + Fleet Scheduler (epic #137)

_best_agent_for_story() went from a simple match to a real 3-stage engine: a weighted capability score (VERIFIER tier, PR review cycles, CI-verification signals), a hard quota-headroom gate across five time windows with a degraded-mode fallback, and a cost tie-break when top scores land within 0.15 of each other. On top of that, synlynk/scheduler.py adds fleet-level batch scheduling — synlynk schedule [--execute] [--max-stories N] — with in-batch headroom accounting so story 5 in a batch of 10 sees stories 1–4's projected spend before it's assigned, and capped retry/reassignment so one flaky story doesn't eat a whole run.

Local Agent — the 5th dispatch agent

local joins claude/codex/agy/grok, dispatching as an aider CLI subprocess against an on-device oMLX OpenAI-compatible endpoint. Zero per-token cost, hardware-gated (an opt-in pytest tier exercises the real integration when hardware is present, cleanly skips otherwise), starting from a conservative capability whitelist that self-widens with verified results. The architecture pivoted mid-design — the original single-shot chat-completion runner couldn't edit files, so it moved to a real CLI-over-local-endpoint shape before implementation began.

Measurement Ledger Hardening (epic #210)

The reason this release got cut in the first place. Fable's 2026-07-12 strategic review named the H0 gate explicitly: "every number synlynk displays is either structurally sourced or visibly labeled as an estimate." Three phases closed it:

Phase 1 — cost_entries gained real provenance columns (cost_source, estimate_basis), a single sanctioned writer (_insert_cost_row(), enforced by a call-site audit test), a rate file (.synlynk/model_rates.json) replacing hardcoded pricing, and a 3-tier t-shirt-size fallback for surfaces with no real token count.
Phase 2 — four structured-output token adapters, one per vendor CLI (Codex, Claude, Agy, Grok), each replacing the old 80/20 heuristic split with a real per-field extraction from that CLI's structured output.
Display layer — Vizor's Effort & Cost tab visually flags estimated-vs-measured rows, and synlynk status gained a RATES line showing when the rate table was last updated (with an explicit ⚠ never updated warning when it's missing).

Every PR in this arc followed the same discipline: dispatched to a rotating real CLI agent, independently verified via git diff against the plan (never trusting the agent's self-report), full-suite regression re-run, cherry-picked, blogged.

Also in this release

Vizor Architect Map v2 replaced a static hand-authored tube-map SVG with a live force-directed graph of workspace repos and typed cross-repo edges, plus an IDE-style file-tree drawer. Vizor Business Goals Panel surfaces goal/story rollups in the HUD. A second __init__.py re-modularization pass extracted 11 focused modules, finishing the transition away from a single-file monolith.

Test Count

Milestone	Tests
v0.11.0 (tag)	868
v0.12.0	1140 passed, 2 skipped

What's Next

Deferred follow-ups from the Measurement Ledger arc, filed as issues and explicitly scoped out of this release: #260 (Savings Ledger), #261 (dispatch-path unification), #262 (surface consolidation), #263 (stage constants staleness). #202 (synlynk jobs stale status display) remains open and tracked separately — not blocking, but unresolved.

The 30-day observation window on synlynk schedule (v1) runs through 2026-08-10 before Fleet Scheduler v2 gets scoped into stories. The bigger open question is what comes after the H0 measurement gate on Fable's roadmap — decide the next horizon's priority before scoping the next named release.