- Repository:
function-master-plot - Local path used during this pass:
/mnt/d/project/函数绘图 - Host used:
Codex - Repository commit during the pass:
a7b85eb - Date:
2026-03-19 - Primary task family exercised:
- read-only repository verification
- plotting test execution
Why this repository was useful:
- it is a real external repository, not the ExperienceEngine repo itself
- it includes an executable read-only plotting test workflow
- its path includes non-ASCII characters, which exposed a host/runtime edge case during validation
At the start of the pass:
ExperienceEnginewas already installed forCodexee doctor codexfrom this repository reported:- adapter installed and enabled
distillation mode = llmsource = explicit_providerevaluation mode = live
- current scope state for this repository:
Scope: scope_05bdb507d983Enabled packs: 0Published packs: 0Compiled targets: 0
Most importantly, this repository had not yet accumulated durable ExperienceEngine history for its own scope:
- no
task_runs - no
outcome_records - no
review_events - no
experience_input_records
This makes it a real cold-start / first-value case rather than a mature repeated-task case.
A real non-interactive Codex host session was executed against this repository with a read-only prompt:
- first run
pwd - then run
npm run test:plotting - report whether the plotting suite passed
The host completed the task successfully.
Observed result from the real session:
pwdreturned/mnt/d/project/函数绘图npm run test:plottingexited with code0- reported checks included:
Plot smoke tests passedImplicit pipeline tests passedKeyboard combo coverage passedWorker eval coverage passedOCR parse coverage passedOCR candidate selection passedPlot performance policy tests passed
During this same pass, Codex logged repeated WebSocket connection failures before falling back to HTTPS transport.
The visible error was:
- UTF-8 conversion failure in the
x-codex-turn-metadataheader - the failing workspace path embedded in that metadata contained the non-ASCII directory name
函数绘图
Practical meaning:
- the host task still succeeded
- but the repository path created a real transport-level friction point for Codex
- this is a host/runtime limitation worth tracking when using ExperienceEngine on repositories with non-ASCII paths
This pass is valuable because of what it did not yet produce.
After the external-repo run, the repository scope still had:
- no
task_runs - no
outcome_records - no
review_events - no
experience_input_records
That means this pass did not yet create a durable ExperienceEngine learning trace for the external repository scope.
The latest ee inspect --last output remained pointed at the most recent ExperienceEngine repository evaluation record rather than this external repository.
This is not a “healthy mature intervention” case. It is a cold-start friction case.
It proves:
Codexcan still complete a real read-only task in this external repository.- A repository path containing non-ASCII characters can introduce real host transport instability before fallback.
- A successful host task does not automatically mean ExperienceEngine has already produced durable scope-local evidence for that repository.
In other words:
- host success is not the same as experience capture success
- cold-start value should be judged separately from wiring success
The most useful part of this pass was diagnostic clarity:
ee doctor codeximmediately showed that the repository was wired but still cold from a Pack/asset perspective- the host run exposed a real environment-level constraint instead of a fake demo success
- the repository is now a good future candidate for measuring first-value improvements once more repeated tasks are run
Two things were clearly friction points:
- non-ASCII repository path triggered Codex WebSocket header encoding failures before fallback
- no durable ExperienceEngine scope-local records were created from this single pass, so the repo remains in an early cold-start state
This is exactly the kind of case that justifies continued work on:
- cold-start visibility
- first-value readiness
- real-host case-study tracking
Current decision for this repository:
- keep using it as an external cold-start validation repo
- prefer repeated similar verification tasks before expecting durable Pack-worthy experience
- treat the non-ASCII path issue as a real host constraint when evaluating Codex behavior
- avoid overstating “successful host execution” as “successful experience learning”
The next useful pass on this repository should:
- repeat the same task family at least one or two more times
- check whether durable
task_runsand input records start appearing for this scope - compare the second and third runs against this cold-start baseline