Skip to content

Releases: ddv1982/flow-opencode

v4.3.9

Choose a tag to compare

@github-actions github-actions released this 08 Jul 21:47

[4.3.9] - 2026-07-08

Candidate accounting coherence lore makes the 4.3.8 orchestration accounting
rules self-consistent, so valid manager records stop bouncing off the schema:

  • decision: "parallel" on implementation-decision records is now rejected
    with one clear message (it stays valid for discovery, audit, and review
    passes); previously every candidateDecision pairing produced contradictory
    errors.
  • Candidate-shaped decisions now require candidate execution evidence on every
    pass kind, so a decision label alone can no longer validate while being
    excluded from candidatePassCount.
  • Candidate and verifier worker counts are checked per subtype instead of
    summed, so one worker may fill both roles on a single pass row.
  • Orchestration pass dedup now remembers every recorded pass id in
    recordedPassIds, so resubmitted completions no longer double-count
    telemetry after the recent-pass window rolls over.
  • The candidate accounting rules are documented once, in
    skills/flow/references/parallel-orchestration.md; the run skill, handoff
    format, and wiki point there instead of restating them, and the two doc
    instructions that contradicted the schema (omitted decision on decision
    records, workerCount=0 with positive candidateWorkerCount) are fixed.

v4.3.8

Choose a tag to compare

@github-actions github-actions released this 08 Jul 18:46

[4.3.8] - 2026-07-08

Parallel orchestration accounting lore makes broad worker use visible without
turning Flow state into a transcript store:

  • flow_feature_complete can now record compact orchestrationPasses for
    serial, skipped, exact-path candidate, isolated-worktree, tournament,
    validation, review, and verifier passes.
  • flow_status reports aggregate pass telemetry under
    session.budget.orchestration, including worker counts, candidate/verifier
    usage, skipped candidate decisions, and recent pass records.
  • Completion now records orchestration telemetry on success, validation-gate
    failures, failed reviews, and needs_input, while deduping retry payloads and
    retaining only the latest compact pass records.
  • Flow planning and running guidance now requires explicit implementation pass
    decisions for broad work and keeps full handoffs, logs, and manager scratch
    artifacts outside .flow/**.
  • README runtime wording now matches the 4.3.6 behavior: completed feature
    counts are telemetry only, not a three-feature stop.

v4.3.7

Choose a tag to compare

@github-actions github-actions released this 08 Jul 13:19

[4.3.7] - 2026-07-08

Package-smoke patience lore keeps the 4.3.6 release path portable across slower
macOS Node 24 runners:

  • The package smoke test now has an explicit timeout large enough for the packed
    consumer declaration check to finish on CI, avoiding a runner-speed-only
    failure after the Release workflow already passed.

v4.3.6

Choose a tag to compare

@github-actions github-actions released this 08 Jul 13:15

[4.3.6] - 2026-07-08

Phase-continuity lore removes the rough stop after three completed features and
adds sharper handoffs around long Flow sessions:

  • Completed feature counts are now telemetry only; approved plans keep moving to
    the next runnable feature instead of requiring a fresh session after three
    completions.
  • flow_status now includes a human-readable progress line, nextFeature,
    pendingFeatures, and remaining feature count so resumed sessions name the
    exact next slice.
  • Hidden worker prompts now fail closed on empty or unstructured handoffs, and
    worker model routing can be configured with OPENCODE_FLOW_*_WORKER_MODEL
    environment variables.
  • Release publishing now has bun run release:monitor, which watches the
    release commit's CI and tag-triggered Release workflow before declaring the
    release healthy.

v4.3.5

Choose a tag to compare

@github-actions github-actions released this 07 Jul 23:34

[4.3.5] - 2026-07-08

Flow now stops long autonomous loops more deliberately without making reviews
shallower:

  • Runtime budget telemetry records completed features since the last phase
    boundary, feature/final review counts, failed review counts, per-feature
    retry attempts, and host token telemetry availability.
  • Failed feature and final reviews now pause by default, allow only one
    autonomous repair plus one retry, then block the feature with a compact
    resume packet instead of continuing to edit in the same root session.
  • Feature completion now records featureReviewDepth and rejects evidence that
    is shallower than the approved feature requires.
  • Large sessions now checkpoint after three completed features and require an
    explicit phaseBoundaryAck in a fresh phase before the next feature starts.
  • Flow planning, running, review, README, maintainer docs, and wiki pages now
    describe scoped review packets, risk-based review depth, retry limits, and
    phase-boundary handoffs.

v4.3.4

Choose a tag to compare

@github-actions github-actions released this 07 Jul 21:05

[4.3.4] - 2026-07-07

Planning and final-review lore sharpened without changing the runtime surface:

  • flow-plan now has a pre-approval quality gate that checks outcome,
    requirements, decisions, uncertainty, feature shape, bounded targets,
    validation levels, dependencies, and review policy before a plan is saved or
    approved.
  • Planning examples now cover bugfix, UI/frontend, runtime/schema, docs-only,
    audit-first, and stronger validation patterns, giving agents better few-shot
    guidance for common Flow sessions.
  • Final review now includes a convergence scan that traces the original goal,
    approved requirements, every planned feature, changed artifacts, and
    validation evidence before returning a passing finalReview.
  • The new planning checklist is shipped through synced skills and bundled
    /flow-plan and /flow-auto command instructions, so fresh and stale skill
    discovery paths get the same guidance.

v4.3.3

Choose a tag to compare

@github-actions github-actions released this 06 Jul 22:56

[4.3.3] - 2026-07-07

Verification only — no behavior change:

  • The live OpenCode smoke test now proves the hidden read-only workers'
    isolation actually binds at runtime: against a real server it asserts each
    worker's resolved permission rules deny the state-changing flow_* tools,
    task, skill, and edit, while keeping flow_status readable. This
    confirms OpenCode compiles Flow's tool-name and wildcard permission keys
    (which are absent from the SDK's simplified permission type) rather than
    silently dropping them, closing the one unverified question from the 4.3.2
    review.

v4.3.2

Choose a tag to compare

@github-actions github-actions released this 06 Jul 22:49

[4.3.2] - 2026-07-07

Correctness and safety hardening from a full-project review, repairing two
edges introduced in 4.3.1 and several older ones:

  • Uninstall no longer deletes a user file that merely resembles a backup name:
    a file counts as removable Flow residue only when its name and its content
    checksum both match Flow's backup format. Doctor applies the same check.
  • Sync no longer overwrites a user-owned skill folder that has files at managed
    paths but no SKILL.md — any managed folder without a Flow marker is left
    untouched, not clobbered without a backup.
  • An empty HOME no longer makes the skills root a current-directory-relative
    path (which sync wrote into and uninstall removed); it falls back to the OS
    home. engines now requires Node >=20.12 (doctor/uninstall rely on it).
  • The CLI no longer aborts on a stray flow-* file in the skills root, and a
    CRLF-converted skill marker is no longer misreported as outdated.
  • Session locks recover from more failure modes: a far-future or foreign-host
    lock timestamp and a recycled process id are now reclaimable instead of
    wedging every Flow call, and stale-lock reclamation no longer races two
    waiters into holding the lock at once.
  • flow_status re-checks the session under the lock before quarantining, so a
    session written by a concurrent process can't be quarantined by mistake; a
    session file with an archive-unsafe id now routes to quarantine recovery
    instead of wedging every archive; and its output is framed as data to blunt
    instruction-injection from a cloned repo's session file.
  • A user command named like an object built-in (e.g. /toString) no longer
    crashes command preflight, and a Flow command invoked with an attachment now
    keeps its worker isolated instead of double-running the instructions.
  • Session history is bounded so long autonomous loops cannot grow the session
    file without limit. Docs corrected to match the current uninstall/doctor
    behavior.

v4.3.1

Choose a tag to compare

@github-actions github-actions released this 06 Jul 10:24

[4.3.1] - 2026-07-06

Doctor and uninstall now account for the .backup files Flow writes when it
retires a locally-edited managed skill file:

  • doctor reports leftover .backup files as action-required instead of
    saying "ok", lists each one, and explains that they hold your earlier edits
    and can be deleted once you no longer need the saved copy.
  • uninstall treats a folder that is pristine apart from Flow-created
    .backup residue as removable and names the removed backup files, so the
    cleanup is no longer silently blocked or silently destructive.

v4.3.0

Choose a tag to compare

@github-actions github-actions released this 06 Jul 10:10

[4.3.0] - 2026-07-06

Parallel orchestration reworked into a single pass playbook, plus sharper
planning and routing:

  • The parallel-orchestration references collapse into one playbook with an
    explicit loop — orient, slice, manifest, fan out, account, verify,
    synthesize, extend or stop — so a manager reads one file instead of four
    cross-linked ones. Every bundled public Flow command got smaller as a result.
  • A new pass manifest doubles as the pre-fan-out coverage gate and the
    accounting contract: one row per slice with its scope, expected coverage, and
    verification tier, and N rows spawned means N handoffs accounted for before
    synthesis.
  • A worker-failure ladder handles a slice that errors, blocks, or returns
    partial: re-spawn once narrower, cover it directly, or carry it into the
    synthesis explicitly as not-covered — never as if coverage were complete.
  • flow-plan gains uncertainty-typed decomposition: resolve specification
    uncertainty by stating an assumption (or asking when a wrong guess is costly)
    and environment uncertainty by inspecting, never by asking.
  • Frontmatter descriptions for flow-plan, flow-run, and flow-review now
    describe only when to reach for each skill, so single-phase asks route
    cleanly.