Repository navigation
[3389 by Claude Opus 5.5] 0.9.0 release logic fix: OpenCode free models are stale, and the failure reason hides the real error. Check current upstream main first. - #3390
Conversation
![propr-dev[bot]](https://github.githubassets.com/images/gravatars/gravatar-user-420.png){kind=avatar}
…there's no Docker in this container, so I couldn't build the agent image or run `scripts/smoke-test-images.sh`. Upstream `main` is the same commit as this branch, so nothing there already covered this. Why the failure reason was wrong. When OpenCode runs the way ProPR runs it (prompt attached as a file), a failed run sends two error events. The first one has your `Model not found: … Did you mean: …` message, then a generic "Unexpected server error" follows. ProPR kept only the last event. On other runs OpenCode sends only the generic event, and the real cause is only in its log, which OpenCode 1.19.5 doesn't write to its log file when the run fails. ProPR's wrapper was also reading logs from the wrong directory. What changed 1. CLI version. OpenCode is now 1.19.5 (the latest release) in `Dockerfile.agent` (both stages), `scripts/build-images.sh`, the core and shared default versions, and the tests that pin the version. `opencode models` and live one-line runs worked without login on all 8 free models. The image-build flag checks also passed against this binary. 2. Free-model catalog. It now lists Big Pickle (still the default), Beam, DeepSeek V4.1 Flash, Flint 40B, GLM 5.3 Flash, MiMo V2.6 Flash, Nemotron 3.5 Lightning and Nemotron 3.5 Super VL. Nemotron 3 Ultra Free is gone: it was listed at first, failed twice with "Model is unavailable", then disappeared from the list. The pricing IDs for Beam and Flint are my best match by name and context size, not confirmed. 3. Retired models. A new table in `packages/shared/src/openCodeCompatibility.ts`, modeled on the Antigravity one, maps each retired model to a current one: - Ling → DeepSeek V4.1 Flash - MiMo V2.5 → MiMo V2.6 Flash - Nemotron 3 Ultra → Nemotron 3.5 Super VL - Muse Spark → Big Pickle - The five older retired models are mapped too. On startup, agents drop retired models and a retired default moves to its replacement. Saved selections, routes, labels, aliases and `/review` picks resolve to the replacement. Any retired ID that still reaches a run is swapped before it is passed to OpenCode, and the health check no longer probes Ling. 4. Failure reasons. - A specific error is no longer replaced by a generic one that comes after it. - The wrapper now passes `--print-logs --log-level ERROR`, and a generic error is replaced by the logged cause with the same `ref`. - The reason still goes through `agentFailureLine` and redaction. I added regressions using the exact event from the issue. Run through the real wrapper and OpenCode 1.19.5, the stored reason is now `Model not found: opencode/ling-3.0-flash-fin-free. Did you mean: deepseek-v4.1-flash-free, glm-5.3-flash-free, mimo-v2.6-flash-free?`. Successful runs still leave stderr empty. One side effect: a failed run's logs no longer include the last 80 lines of OpenCode's log file, but that step was reading the wrong directory and never worked. 5. Changelog and docs. 0.9.0 has Changed bullets for the CLI and the catalog, the Retired models entry is updated, and there's a Fixed bullet for the failure reason. The OpenCode model table in the docs is updated. Test results - A run of 95 test files covering the changed code passed, as did the OpenCode agent, parser and model-definition tests (including `test/modelDefinitions.test.ts`). After that run I made one small fix to how logged messages are unescaped and reran the OpenCode tests, which pass. - Root lint and typecheck, workspace lint with `--max-warnings 0`, typechecks for api, core, shared and the UI, and `EXPECTED_VERSION=0.9.0 RELEASE_TAG=v0.9.0 RELEASE_CANDIDATE=true npm run release:verify` all pass. - The optional live native-log test passes against the real 1.19.5 binary. - `test/modelAliases.test.ts` always reports a pass because it ends with `process.exit(0)`, which hides 15 failures that already exist on `main`. With that line removed, my new tests pass and the 15 failures are the same ones as on `main`. I captured the new OpenCode model list in the agent settings at desktop and mobile width as previews. Nothing was merged, tagged, published or deployed.
|
/review |
|
✅ AI Code Review Complete requested by @integry Posted 1 review: |
🔍 AI Code Review — codex:gpt-6-astraOverall EvaluationThe PR refreshes OpenCode versions and models, adds compatibility routes, and improves failure reporting. It needs a localized correction before merge: switching a saved plan away from a retired model can leave its old routing label active. ✅ Preserves useful errors — A specific JSON error survives a later generic server error, with regression coverage for both event orders. ✅ Correlates logged causes — Generic errors are replaced only by stderr causes matching their error reference; unrelated references are tested. ✅ Consistent version updates — Docker stages, build defaults, shared defaults, and core defaults agree on OpenCode 1.19.5. This review uses static analysis of the supplied code; no commands were run. The issue specification is truncated, so unseen acceptance criteria could not be assessed. Current checks report no failures and three pending checks. Merge blockersEvery finding below was introduced by this PR and must be resolved before merging. F1: 🔴 Remove the original retired routing label
SuggestionsThese are optional follow-ups and are not sent to S1: 🟢 Exercise the wrapper behaviorThe wrapper tests inspect script text. An executable test using a stub ScoreThe error-reporting changes and compatibility tests are useful, but the persisted-label regression requires correction. No current checks have failed; three remain pending. Score: 6/10 🤖 Review Details
Copy and edit to choose records from this review: 🤖 Review by ProPR |
|
/fix F1 |
|
✅ Applied the requested follow-up changes in commit 0493b39
Summary of ChangesI fixed F1. When you change an issue's model selection away from a retired OpenCode model, the old label that is actually on the issue (like The change
The same defect elsewhere Tests
One side effect: changing away from a retired model now sends one or two extra label-removal calls to GitHub. A label that isn't on the issue comes back as "not found", which is treated as success, so nothing fails. There's nothing visual to preview. 🤖 Implementation Details
Undo Changes • View Task Execution 💡 ProPR Slash Commands
🤖 Generated by ProPR |
… from a retired OpenCode model, the old label that is actually on the issue (like `llm-opencode-muse-spark13-free`) now gets removed, along with the newer replacement label. Before, only the replacement label was removed, so the old one stayed on the issue and started an extra Big Pickle run. I fixed F1. When you change an issue's model selection away from a retired OpenCode model, the old label that is actually on the issue (like `llm-opencode-muse-spark13-free`) now gets removed, along with the newer replacement label. Before, only the replacement label was removed, so the old one stayed on the issue and started an extra Big Pickle run. **The change** - `packages/shared/src/openCodeCompatibility.ts`: a new `retiredOpenCodeModelLabel(id)` gives back the label a retired model had before it was retired. It accepts the `opencode:` and `opencode/` forms. - `packages/api/routes/planIssueHelpers.ts`: a new `getAttachedLlmLabels(model, alias)` returns every label an issue could already have for its saved model: the current label and, for a retired model, the old label too (including the agent-alias version when an alias is set). `handleMultiAgentImplementation` now removes all of these. Any label the user still has selected is left alone, so picking a retired model's replacement keeps the replacement label and only drops the old one. **The same defect elsewhere** `syncModelLabels` in `packages/api/routes/planIssueConfigSync.ts`, which runs when one issue's model or agent is edited, also built its removal list from the replacement label only. I fixed it the same way. The queued-issue repair path doesn't have this problem, as the review said. **Tests** - New `test/planIssueRetiredLlmLabels.test.ts` runs the real `handleMultiAgentImplementation`, with only the GitHub client, queue and database replaced by fakes. It covers an issue that already has the retired label, both when switching to a different model and when switching to the retired model's replacement. - With the old removal logic put back temporarily, both handler tests fail. With the fix, they pass. - `test/planIssueConfigSync.test.ts` gets a matching test for the single-issue edit path. - I added the new export to the test mocks in `test/planIssueConfigSync.test.ts`, `test/planIssueHandlers.followup.test.ts` and `packages/api/test/mcpEpicDispatch.test.ts`. Without it the config-sync module wouldn't load in two of them, and `mcpEpicDispatch` would quietly call the real label lookup. - Passing: the new file (3), `planIssueConfigSync` (4), `planIssueHandlers.followup` (9), `planIssueHelpers` (4), `mcpEpicDispatch` (72) and `planIssueImplementationTrigger` (2). - Typechecks for shared, api and the root pass, and ESLint with `--max-warnings 0` is clean on the changed api files. The shared source and root `test/` files aren't covered by any lint config. One side effect: changing away from a retired model now sends one or two extra label-removal calls to GitHub. A label that isn't on the issue comes back as "not found", which is treated as success, so nothing fails. There's nothing visual to preview. PR: #3390 Comment by: @integry (ID: 6506967925) Model: Claude Opus 5.5
|
/review |
|
✅ AI Code Review Complete requested by @integry Posted 1 review: |
🔍 AI Code Review — codex:gpt-6-astraOverall EvaluationThe PR updates OpenCode’s bundled version and free-model catalog, preserves retired selections through compatibility routes, and improves failure reporting. It appears merge-ready within the supplied scope, conditional on the pending Validate Changes check. ✅ Specific failure preservation — The parser retains a specific error when a generic server error follows, and stderr recovery requires the matching error reference. ✅ Consistent compatibility routing — Retired models resolve to current models across execution, saved-selection validation, token limits, and label generation. ✅ Historical label cleanup — Configuration changes and multi-agent implementation remove historical retired-model labels, with regression coverage. This was a static review; no commands were run. The original issue description is explicitly truncated, so acceptance requirements beyond the visible text could not be assessed. Merge blockersNo merge blockers. SuggestionsThese are optional follow-ups and are not sent to S2: 🟢 Test configured-agent historical labelsAdd coverage for ScoreThe supplied changes address the visible failure and migration paths with focused regression tests. No verified blocker was found; authoritative checks report 27 passed, none failed, and one pending. Score: 9/10 🤖 Review Details
Copy and edit to choose records from this review: 🤖 Review by ProPR |
AI Implementation Summary
Closes #3389
Branch:
3389/claude-opus-5-5-0-9-0-release-logic-fix-o-20261029-1836-0b0Commits: ✅ Changes committed (085291a)
AI Processing Completed
Execution Details:
Summary:
I've fixed all five points of #3389. One requested check didn't run: there's no Docker in this container, so I couldn't build the agent image or run
scripts/smoke-test-images.sh. Upstreammainis the same commit as this branch, so nothing there already covered this.Why the failure reason was wrong. When OpenCode runs the way ProPR runs it (prompt attached as a file), a failed run sends two error events. The first one has your
Model not found: … Did you mean: …message, then a generic "Unexpected server error" follows. ProPR kept only the last event. On other runs OpenCode sends only the generic event, and the real cause is only in its log, which OpenCode 1.19.5 doesn't write to its log file when the run fails. ProPR's wrapper was also reading logs from the wrong directory.What changed
CLI version. OpenCode is now 1.19.5 (the latest release) in
Dockerfile.agent(both stages),scripts/build-images.sh, the core and shared default versions, and the tests that pin the version.opencode modelsand live one-line runs worked without login on all 8 free models. The image-build flag checks also passed against this binary.Free-model catalog. It now lists Big Pickle (still the default), Beam, DeepSeek V4.1 Flash, Flint 40B, GLM 5.3 Flash, MiMo V2.6 Flash, Nemotron 3.5 Lightning and Nemotron 3.5 Super VL. Nemotron 3 Ultra Free is gone: it was listed at first, failed twice with "Model is unavailable", then disappeared from the list. The pricing IDs for Beam and Flint are my best match by name and context size, not confirmed.
Retired models. A new table in
packages/shared/src/openCodeCompatibility.ts, modeled on the Antigravity one, maps each retired model to a current one:On startup, agents drop retired models and a retired default moves to its replacement. Saved selections, routes, labels, aliases and
/reviewpicks resolve to the replacement. Any retired ID that still reaches a run is swapped before it is passed to OpenCode, and the health check no longer probes Ling.Failure reasons.
--print-logs --log-level ERROR, and a generic error is replaced by the logged cause with the sameref.agentFailureLineand redaction.I added regressions using the exact event from the issue. Run through the real wrapper and OpenCode 1.19.5, the stored reason is now
Model not found: opencode/ling-3.0-flash-fin-free. Did you mean: deepseek-v4.1-flash-free, glm-5.3-flash-free, mimo-v2.6-flash-free?. Successful runs still leave stderr empty. One side effect: a failed run's logs no longer include the last 80 lines of OpenCode's log file, but that step was reading the wrong directory and never worked.Changelog and docs. 0.9.0 has Changed bullets for the CLI and the catalog, the Retired models entry is updated, and there's a Fixed bullet for the failure reason. The OpenCode model table in the docs is updated.
Test results
test/modelDefinitions.test.ts). After that run I made one small fix to how logged messages are unescaped and reran the OpenCode tests, which pass.--max-warnings 0, typechecks for api, core, shared and the UI, andEXPECTED_VERSION=0.9.0 RELEASE_TAG=v0.9.0 RELEASE_CANDIDATE=true npm run release:verifyall pass.test/modelAliases.test.tsalways reports a pass because it ends withprocess.exit(0), which hides 15 failures that already exist onmain. With that line removed, my new tests pass and the 15 failures are the same ones as onmain.I captured the new OpenCode model list in the agent settings at desktop and mobile width as previews. Nothing was merged, tagged, published or deployed.
Detailed Logs:
a8401383-b1fd-419d-aaec-19621b10c196Log files stored at:
/usr/src/app/logs/executions/issue-3389-2026-10-29T19-30-37-405Z-conversation.json/usr/src/app/logs/executions/issue-3389-2026-10-29T19-30-37-405Z-output.txtLatest Conversation Messages
This PR was created automatically by ProPR after processing issue #3389.
💡 Need changes?
Comment on this PR to request refinements — the AI agent monitors comments and will update the implementation based on your feedback. Keep iterating until you're satisfied!
Visual preview
OpenCode free models (desktop)
AI Agents configuration: the OpenCode card lists the eight free models OpenCode CLI 1.19.5 serves; the retired Ling, MiMo V2.5, Muse Spark and Nemotron 3 Ultra entries are gone
Original unavailable: managed storage is unavailable or the upload has expired.
OpenCode free models (mobile)
The same OpenCode model list at 390px width
Original unavailable: managed storage is unavailable or the upload has expired.