release: 0.6.7

[xdotli]

Cuts 0.6.7 from current main. Everything in this release is backward-compatible; the BREAKING environment→sandbox task-config rename (#966) is deliberately not here — it opens 0.7 and merges after this tags.

Contents — 19 PRs (#949–#967)

Highlights, all dogfood-driven (7 blind fresh-user rounds this cycle, each validating the previous fix or surfacing the next):

• Console you can actually watch. Live per-task activity in the dashboard (38 calls · 412.0k tok), phase labels so the cell is never blank, a ~45s heartbeat for non-TTY runs, and live token totals mid-run — including mid-prompt via the gateway's existing capture loop, which is what single-task evals actually need (fix(eval): console UX from fresh-user dogfood — progress heartbeat + clean teardown #951, feat(cli): live per-task activity in the eval dashboard; console polish #956, feat(cli): phase labels in the activity column; one-line failure reasons #957, feat(cli): live token usage in the dashboard footer; token-forward activity cell for single-tool agents #963, feat(cli): mid-prompt live token usage via the gateway's existing live capture #965).
• Console you can act on. Fractional rewards everywhere the counts binarize (mean reward 0.30), one failure-reason line per failed task mined from in-memory metrics or bounded verifier artifacts (CTRF / reward.json / stdout tail), multi-failure roll-ups, and a (details: …/verifier) pointer whenever artifacts exist (feat(cli): failure reasons from verifier artifacts (CTRF / test-stdout) #959, feat(cli): surface fractional rewards in console summaries #960, feat(cli): full failure count on per-task console lines; preserve param ids #962, feat(cli): env0-shaped failure breakdowns; details pointer keyed off on-disk artifacts #964).
• Zero-config environments. The env0 pins ship inside the wheel, so --environment-manifest env0@prod resolves on a bare pip install — verified from an empty directory by an independent fresh-user run (feat(environment): built-in manifest registry shipped in the wheel #961).
• Reliability fixes with teeth. A host-side hard rollout deadline for wedges below the phase watchdogs, and the PTY transport bug where an agent's initialize response glued to a shell prompt was silently discarded — a task that failed 7/7 at any timeout now completes at the default 60s (fix(rollout): hard deadline around the rollout lifecycle + verifier zero-output-timeout retry #949, fix(acp): decode protocol JSON glued to PTY shell noise; env-configurable handshake timeout #958).
• Harbor 1.3 [task] version acceptance, clawsbench task.md migration, env0 pin sync + restored outage perturbation, compact flow-style frontmatter arrays (fix(task): accept Harbor 1.3 [task] version field #953, chore(clawsbench): migrate to task.md + actionable no-manifest verifier error #952, fix(env0): sync drifted service pins + fail loud on zero startable services #954, docs: fix post-merge drift from #949–#954; restore env0@outage perturbation #955, feat(task): compact flow-style arrays in emitted task.md frontmatter #967).

Release-specific changes in this PR

• Version 0.6.7 (from .dev0), CITATION date, CHANGELOG 0.6.7 section with per-PR entries.
• env0 pins repointed to the org-owned ghcr.io/benchflow-ai/env0:0.2.0: the wheel-shipped pins named a personal Docker Hub image while their own comments and docs/environment-plane.md declared the ghcr image authoritative, and base_image is recorded as rollout provenance — so every env0 result carried the wrong base. Validated on Daytona before flipping: env0 auth-least-privilege-summary, 8/8 services ready, reward 1.0 in 2.8 min. PR-archaeology comments stripped from the files that ship to PyPI.
• Docs: mean_reward documented in the summary.json layout; the failure block and live-token footer documented in the CLI reference; the external-agents cost figure softened (the underlying rate table self-describes as a placeholder).

Gates

ruff format/check, ty — clean. Full suite: 5368 passed (3 known host-specific failures: the rbenv-sensitive heartbeat test and two smoke-wiring credential tests that trip when real credentials exist on the host).

Post-merge steps for the maintainer: tag v0.6.7 on the merge commit and publish to PyPI per the release-manager process.