0.4.0

Paid free-tier failover (first paid go_ranked row while the latch is fresh), auto mode with routing sync, opencode/auto virtual model, host detection, /modelselect command, Agent/Think status fields, rebuilt Work Status bundle. See CHANGELOG.md for the full list. Use dianlight/opencode-modelselect-action@0.4.0 (or @v1).

0.3.0

[0.3.0]

Fixed

• tier: auto now honors auto-preference: the free probe strips the
  opencode/ engine prefix before calling Zen (prefixed names answer 401
  "not supported" even for valid keys) and probes the chat and responses
  endpoints in parallel (new probe-responses-url input, derived from
  probe-url by default) since free models live on either per the Zen docs
  endpoint table. A 400 session gate (MissingSessionID / "only be used in
  OpenCode") counts as selectable — the key is accepted and free serves the
  downstream OpenCode step — so free-first picks free even while Go quota
  remains. A 401 from one tier falls back to the other; the step fails as
  invalid token only when both tiers reject auth

Removed

• No-op file-sync job, keep deprecated-workflow cleanup

Changed

• tier input is now dynamic when omitted (was go): auto when a token
  (opencode-token or OPENCODE_API_KEY) is available, else free
• README Inputs table now lists Description before Default

Added

• README badges (release, last commit, issues, pull requests, license),
  Sponsor section (copied from dianlight/srat, incl. OpenCode Go referral),
  and License section
• README Task types cross-links from the task-type input, requirements, and
  How it works; LiveBench Score Reference now explains the Value column
  (Overall ÷ Blended $/1M) and what a missing — means (Free $0 cost or
  unknown pricing)

0.2.0

Removed

• Workflow Model Audit: models are resolved dynamically at runtime from the
  central data/model-config.json (committed directly each run), so the
  per-workflow audit table, the issue's suboptimal-configurations section,
  the workflow scanner (scan_workflows, classify_task_type,
  classify_model_status), data/workflow_scan.json,
  config/workflow-task-map.yaml, and the signals keys in
  config/task-types.yaml are gone; the maintenance issue now only covers
  model coverage (config/model-scores.yaml PRs)

Added

• Model Audit issue: new "Correct the model coverage issues" checkbox that opens
  a PR scoped to config/model-scores.yaml only (removes stale fallback entries
  now on LiveBench, adds scores for models missing data)
• Test coverage: new test/coverage-extra.test.js (36 tests, JS at 99% lines
  via npm run coverage) and tests/test_maintenance.py (52 offline unit
  tests for the maintenance script via mise run test-python); CI lints
  tests/ and runs the Python suite

Changed

• Model evaluation now uses a blended in/out token-cost selector: each tier
  picks the cheapest blended $/1M cost among models within the free-first
  threshold of the top LiveBench score (weights from cost_blend in
  config/task-types.yaml, default 75% in / 25% out; unknown costs sort last)
• Rename the project and repository to Opencode Modelselect
  (dianlight/opencode-modelselect-action): update action.yml name,
  config-url default, README, workflows, package.json, and User-Agent

Added

• New select-model GitHub Action (action.yml + src/index.js, Node 24, zero
  dependencies): preselect the OpenCode model for a task-type + tier
  (go/free) step before the OpenCode step, reading the live central
  data/model-config.json. Fails hard when the config is unreachable or the
  task-type has no entry (optional fallback-model escape hatch). Downstream
  usage: dianlight/opencode-modelselect-action@<tag> with task-type/tier inputs,
  then model: ${{ steps.resolve.outputs.model }} in the OpenCode step
• Key the central data/model-config.json by task-type only
  (task-types.<name>.{go,free}) for all 11 task types, so any downstream
  workflow can preselect a model without workflow/job coupling
• sync-actions now opens (or updates) a repo-sync/cleanup-deprecated-opencode-workflows
  PR in each target repo deleting the six deprecated no-op workflows (opencode.yml,
  opencode-triage*.yaml, opencode-implement.yaml, opencode-review.yaml)
• Fetch Zen model prices from the Zen docs pricing page during maintenance and
  store them per model in data/zen_models.json; treat models published as
  "Free" (e.g. big-pickle, which has no -free suffix) as usable free models
• Surface in/out token costs ($/1M) in the README recommendation/audit cells
  and the score reference table (new In/Out/Blended/Value columns), include
  input_cost/output_cost/blended_cost in benchmark_results.json and
  audit_results.json, and report paid models with unknown pricing under
  coverage_issues.json missing_prices (warn-only)
• New select-model max-cost input (blended $/1M budget cap): an over-budget
  pick is replaced by the best-scoring ranked model within budget from the new
  per-task-type go_ranked/free_ranked lists (best-to-worst with scores and
  costs) in data/model-config.json; the step fails when nothing fits unless
  fallback-model is given. New model-cost output reports the resolved
  model's blended $/1M
• New select-model tier: auto mode: probes live quota with opencode-token
  (or OPENCODE_API_KEY) via a tiny free-model request plus GET /zen/go/v1/usage for the Go plan windows. auto-preference
  (free-first/go-first) sets the order, max-wait-seconds /
  poll-interval-seconds poll until quota frees up instead of failing fast.
  New tier-selected output reports the tier actually used

Removed

• Delete the 6-process pipeline workflows (opencode-pr-review.yml,
  opencode-pr-comment.yml, opencode-issue-handler.yml; opencode-implement.yaml
  was already gone): drop their entries from .github/sync.yml and
  config/workflow-task-map.yaml, replace WORKFLOWS.md with a deprecation
  pointer, and extend the sync-actions cleanup job to open downstream removal PRs
• Delete .github/scripts/resolve-model.sh; model resolution now lives in the
  select-model action. All active workflows use it (uses: ./ here,
  dianlight/opencode-modelselect-action@<tag> downstream) and it is dropped from
  .github/sync.yml since model selection needs no file sync
• Drop the per-workflow fix-suboptimal-configs checkbox flow from maintenance
  issues: workflows no longer pin models, so apply-model-config is the only
  model fix path
• Drop the six deprecated no-op workflow entries from .github/sync.yml and delete
  the local stub files; deletion in target repos is now handled by the sync-actions cleanup PR job
• Drop the alt models concept and its token-multiplier-driven second
  checkbox from maintenance issues, and remove token_multipliers from
  config/model-scores.yaml; the maintenance issue now offers only the single
  "Apply the proposed model config update" checkbox
• Remove the last /oc / /ocf slash-command remains: delete
  .github/scripts/auth.sh (command parser, the only file still synced
  downstream — .github/sync.yml now carries an empty file list), RUNBOOK.md,
  and .github/workflows/WORKFLOWS.md; drop the dead conditional-model
  expression parser from scripts/opencode_maintenance.py (all steps resolve
  via the select-model action now, audit shows the free model as free instead
  of /ocf); reword action.yml, README.md, AGENTS.md, and the maintenance
  workflow to go/free tiers with no slash-command references; extend the
  sync-actions cleanup job to delete auth.sh downstream as well

Fixed

• Restore emoji icons (✅/📋/❌) in the README LiveBench Score Reference
  Source column, replacing the plain v/f/x letters
• Fix Handle Checked Tasks failing with Duplicate header: Authorization by
  passing use_github_token: true to the OpenCode action (the OIDC token
  exchange previously added a second Authorization header on top of the one
  persisted by actions/checkout); checkout now also fetches full history and
  the job grants pull-requests: write so the OpenCode CLI can open the PR it
  wraps task changes in (previously 403 Resource not accessible by integration)
  after the task itself succeeded
• Fix the opencode-pr-comment workflow failing with
  fatal: could not read Username for 'https://github.com' during the OpenCode
  step's git push (process-3): with use_github_token: true the OpenCode CLI
  skips configuring git credentials, so push-capable jobs must persist the
  actions/checkout token and grant contents: write. process-3 now persists
  credentials, configures the git identity, and grants contents: write;
  process-6 also grants contents: write for its push path

Resolves #31

Release 0.1.3

Full Changelog: 0.1.2...0.1.3

Release 0.1.2

Full Changelog: 0.1.1...0.1.2

Release 0.1.1

What's Changed

• Update actions/checkout action to v7 by @renovate[bot] in #19
• big-pickle scores added; issue #20 closed by @github-actions[bot] in #22
• Add qwen3.8-max benchmark scores to model-scores.yaml by @github-actions[bot] in #24
• Update central model config by @github-actions[bot] in #28
• 🐛 fix(ci): apply all 10 coderabbit findings from #29 by @dianlight in #30

Full Changelog: 0.1.0...0.1.1

0.1.0

First functional release