general-agent-experience.md

General Agent experience: command and file workflows

This iteration keeps ChatGPT's web client as the reasoning caller. WebCodex owns the authorized tools and evidence; it does not add another model service or a second Agent loop. Research and local acceptance were performed on 2026-10-01.

Research and current implementation

Primary source	Relevant practice	Application to WebCodex
Codex execution handler	Distinguish live execution identity, completion and bounded output.	Retain WebCodex's original Job rather than redispatching; distinguish readiness waiting from log observation.
SWE-agent ACI	Design file viewers, search and command feedback for model decisions; check edited code.	Reuse bounded reads/search and guarded edits, and validate the task's actual effects with focused checks.
Aider coder	Feed edit/lint/test failures into bounded repair cycles.	Make diagnosis, fresh observation, repair and validation explicit to the web caller; do not automatically replay uncertain effects.

The inspected upstream 78826899b58d1284906db92e1b4287956c5b3e69 already contains #812: MCP failure-log deduplication, output metadata/closeout guards, Work Result retention and general task guidance. It also already has revision-fenced edits, transactional preflight, same-Job continuation and bounded readiness waits. These mechanisms were reused.

The remaining demonstrated gap was in default compact discovery: its bootstrap still said "coding/review", and command entry descriptions omitted the independent-work/readiness-wait choice. The fix updates the legacy and stateless bootstrap, editor, process, shell and script selection descriptions. Every changed description fits the existing 420-character bound with its final recovery clause intact. Business input schemas, language choices, annotations, authorization and execution semantics stay canonical. Effects on model decisions remain an inference until real model runs are collected.

Local acceptance and reproducibility

The latest upstream source was merged into fork main without conflicts. The result 3e4fcda5cbc5e4f85b479cd789af6a16f8f15892 has the same source tree as 78826899; its merge preserves the fork's history. The unrelated local issue-784-continuity commit remains on its own branch. Only synchronized main was pushed; the new improvements remain on codex/general-agent-experience.

Use the same current harness source for both current-contract runs, with --profile upgraded for both. The historical baseline profile belongs to the pre-output-receipt contract, not to this comparison. Reports retain each actual build identity, binary hashes and the common fixture-script fingerprint. Setup and upgrade-only checks are separate from the six comparable scenarios.

The driver uses disposable loopback Server/Runner processes and temporary Projects. Locale, timezone, Python environment and the fixture Git commit date are fixed. It independently checks actual files/bytes/hashes and CSV statistics, observes expected exit failures, verifies one original pending Job through its captured call sequence, and checks stale rejection before fresh-read recovery. Missing-output and large-output/Work Result checks run separately. The App client is an explicit synthetic fixture; it does not prove real ChatGPT card rendering.

cargo test -p webcodex --lib mcp::tests::tools -j 2
cargo test -p webcodex --lib mcp::tests::execution_feedback -j 2
python3 -m unittest discover -s scripts/tests -p test_generic_agent_acceptance.py
python3 -m unittest discover -s scripts/tests -p test_agent_loop_report.py
cargo build --profile dogfood -p webcodex --bin webcodex-server -p webcodex-runner --bin webcodex-runner -j 2
python3 scripts/e2e_generic_agent_ws.py --bin-dir target/dogfood --profile upgraded --source-version YOUR_BUILD_COMMIT --report /tmp/webcodex-generic-upgraded.json

Observed local results

Both binary pairs reported version 0.4.4 with dirty=false. The baseline build commit was 3e4fcda5cbc5e4f85b479cd789af6a16f8f15892; the improved build was 5c49bcb8bde148e7bdac7515896a5098b845ac38. Documentation added after these builds is not part of either binary's runtime source. The shared harness SHA-256 was 4e80f56c02eb0159b3537ae07f486a0663853c120be77e0c34cfdec800b0c253.

Comparable evidence	Latest-upstream baseline	Improved
Independent task scenarios passed	6/6	6/6
Separate missing/large output checks passed	2/2	2/2
Meaningful outer calls	23	23
Actual MCP response bytes	43,011	43,011
Actual request bytes	9,418	9,418
Expected unsuccessful tool calls	3	3
Comparable task span, seconds	1.236869	1.262349

All six independently verified scenario evidence objects agree after excluding the ephemeral Job identity. The three unsuccessful tool calls are the intentional exit-7 command, broken-code assertion and stale edit. Both missing-output guards also correctly returned a blocking outcome in their separate checks.

The scripted sequence uses the same calls irrespective of selection descriptions, so these equal counts/bytes do not measure a model-selection improvement. The single-run timing difference is descriptive and does not establish a speed regression or improvement. The changed discovery behavior is covered separately by 53 MCP tool tests, including compact/full business-schema parity and legacy/ stateless instruction routing; four execution-feedback tests and 79 Python tests also passed. Final Rust formatting and diff whitespace checks passed. Both production builds emitted the same pre-existing unused Value import warning in src/tool_runtime/trace_diagnostics.rs; it was outside this change.

Raw local reports retain the binary hashes, actual call sizes/order and scenario evidence. They are saved under ignored target/agent-experience-20261001-5c49bcb8/{baseline,improved}.json in the feature worktree; no credential headers or full request/result payloads are recorded.

Real ChatGPT replay

The six generic_* entries in scripts/agent_loop_cases.json provide the task prompts, fixture recipes, expected files and independent correctness gates. Use the existing Runner Project registration workflow for an isolated prepared directory. Prepare only the selected case's inputs from the named fixture method; the deterministic driver is not a recorded model run.

For each comparison, use a fresh identical directory, exact fixture Git base, same prompt, public model selection and caller surface, and record the runtime build separately. Use Direct and Host Code Mode as the existing protocol specifies; experimental Code Mode can only invoke admitted nested tools, and Job operations remain outside it. The stale-edit case requires the fixture operator to append its sentinel after the first model read and confirm the write before the stale attempt. A missing-output case is an expected negative guard check.

Collect existing ActionAudit metrics with scripts/agent_loop_report.py and record manually observed contract repairs using its bounded annotation. Expected command/test/stale failures are ordinary business evidence, not contract repairs. Do not infer private reasoning, model turns or token use from service gaps or response bytes. The reporter's surface comparator is for Direct versus Code Mode; same-surface runtime-version reports can be read side by side without claiming that comparator's pair gate passed.

No real ChatGPT benchmark or existing-service deployment was performed in this iteration. The local scripted run demonstrates runtime contracts and preservation of effects. It cannot establish that the model understands every task better or will always continue in another turn. Those conclusions require actual web-client samples, including cached-tool discovery and observed task completion.