CoTrading DocsTroubleshoot skill discovery and failed tasks
Find the first failed stage, retain evidence, and recover without overwriting skills, input files or previous results.
Diagnose the first failed stage. Download, skill discovery, readable inputs and completed computation are separate outcomes. A successful earlier stage cannot certify the later ones.
Failure map
| Symptom | Check first | Safe next step |
|---|---|---|
| SHA-256 differs | Selected host/version and actual ZIP | Stop; obtain the matching release |
| Skill absent | Folder nesting, SKILL.md and name | Correct selected folder; refresh session |
| Same name exists | Manifest version and local edits | Preserve backup outside scanned directories |
| Python unavailable, including text methods | Python 3.12 for S01/S02/S04/S05 | Stop fixed-tool execution; arrange documented isolated setup |
| PDF unreadable | Extracted numbers and column order | Supply permitted readable text |
| Missing data/rule confirmation | Manifest, calendar, actual confirmation | needs-input; list exact missing item |
| Model quota/network failure | Host error | Preserve inputs and incomplete status |
| Different numeric result | First divergent daily/event row | Compare rule/data/version and timing |
The official discovery mechanisms are documented by OpenAI and Claude Code, checked 2026-10-30. They do not replace package-specific verification.
Discovery example
For S01, Codex's key file is %USERPROFILE%\.agents\skills\cotrading-earnings-review\SKILL.md; Claude Code uses %USERPROFILE%\.claude\skills\cotrading-earnings-review\SKILL.md. If you see an extra version or archive directory, the selected folder may be at the wrong level. If the destination already exists, do not overwrite until its version and edits are preserved.
Use the conversation's actual selector. A slash or dollar skill mention pasted into PowerShell is not a skill invocation. Restart the session if the current documented reload behavior has not picked up an update.
Data and dependency example
The four-day synthetic ETF sample cannot run 63-interval momentum: it needs 64 complete observations. Missing ETF_B on one date should produce a precise gap and stop. Forward-filling or silently dropping it changes the experiment and is not a repair.
The current 0.1.0 A-batch packages S01/S02/S04/S05 all require Python 3.12, including the earnings, thesis and specification text tasks. Their bundled deterministic tools use only the standard library and do not access the network; this does not make remote model processing offline. py -3.12 --version checks availability only. Record the executable/environment and error before installing anything. If missing or incompatible, stop the fixed-tool task, preserve inputs and prepare the documented isolated environment separately. Do not use model mental arithmetic as evidence that the tool ran. Do not install third-party Python packages for these teaching tools, change global settings, bypass security settings, elevate privileges automatically or run an unrelated bulk installer. OCR or other additional parsers have their own requirements.
Preserve honest run states
Use a new run directory for every attempt. needs-input means something must be supplied or confirmed; failed means execution/tool failure; completed describes the actually completed limited scope. Do not show an old report as the latest attempt or fabricate output after a failed calculation. A partially complete review must identify its unanswered questions.
Feedback, updates and removal
Provide method version, host/version, platform, failure stage and a redacted error. Do not send keys, holdings or full research documents by default. Updating is an explicit version choice; retain local edits and old outputs. Remove only the selected installed folder when uninstalling, never the separate research folder or other skills.
For a clean first task return to Codex CLI or Claude Code. Free packages do not imply free model quota or paid-source access.