[xizhuomengcontin]

Adds OrcaReplay to coding, following the four steps in CONTRIBUTING.md.

Disclosure: I maintain OrcaReplay. Apache-2.0, Node 20+.

Why it belongs in coding rather than a new category

The coding category already holds the agents themselves — Claude Code, Codex, Cline, goose, Pi — plus tooling around them (agent-skills, Taste-Skill). This is tooling of the second kind: a way to run a past session again.

It records the agent's exchange with its model provider at the HTTP boundary, so the recording is the bytes the agent actually sent and received rather than a summary it wrote about itself. orca replay last runs that exchange again with no provider contacted and nothing spent.

orca replay last --from <n> --model <other> replays the prefix from the recording and then continues on a different model, so the model is the only variable. This is bounded by checkpoints, not by arbitrary events: deriveCheckpoints only produces a checkpoint where there is both a complete conversation prefix and a filesystem snapshot, --fork refuses a run that has none, and a --from that does not land on a checkpoint is snapped back to the nearest prior one with fork.snapped. So a reader comparing models from a particular event starts from an earlier state, or finds the run cannot fork.

Against the inclusion criteria

• Relevant — agent coding / agent infrastructure. Several coding entries (Claude Code, Codex, Cline, goose) are agents it records directly; goose and Claude Code have adapters and Claude Code is validated against a real bug fix.
• Active — commits this week.
• Open source — Apache-2.0. The on-disk trace format is specified separately under CC BY 4.0 (spec/orca-trace-v0.md), so a recording is readable without the CLI.
• Documented — README with install and usage; orca quickstart writes a small buggy project plus a recording and replays it, so the claim is checkable in one command.
• Traction — published on npm as orcareplay. (Star count left to update-stars.js rather than written into the PR text.)

Not a cross-list entry — it is not a multi-agent framework.

Steps followed

node scripts/generate-readme.js   -> regenerated
node scripts/validate.js          -> Validated 121 projects across 10 categories / Validation passed.
git diff --check                  -> clean
node scripts/check-links.js       -> Results: 121 ok, 0 failures, 0 warnings

Run on Node 22.23.2. The branch is rebuilt on current main, so the earlier conflict is gone; the diff is +13 −0 in data/projects.json and +1 −0 in the generated README.

What I am not claiming

• A matching replay is not a determinism result. It proves the recorded exchange reproduces, not that the agent is deterministic or that a fresh run would fail the same way.
• Replay withholds the model provider only. It is not a sandbox, and it is not offline — recorded tool calls still execute for real, and a tool that opens its own connection still reaches the network.
• Fork is bounded by available checkpoints, as described above.
• The default adapter does not capture every call; embedding calls, for one, are not captured.

🤖 Generated with Claude Code

[autogame-17]

Reviewed current head f0808deb762d2f8bf4cc0eeaffbf5b9adae4d797, including the canonical repository, license and replay implementation. The coding-tool scope is appropriate; please tighten one factual claim before inclusion.

data/projects.json:1543 and generated README.md:156: “replay can fork from any step” should say that it forks from available derived checkpoints. At pinned upstream 6af80889:

• deriveCheckpoints requires both a complete conversation prefix and a filesystem snapshot.
• The fork command rejects runs without checkpoints and snaps a non-checkpoint --from sequence back to the nearest prior checkpoint, emitting fork.snapped.

A reader comparing models from a particular event can therefore start from an earlier state, or find that the run cannot fork. “Replay can fork from available checkpoints onto a different model” preserves the useful capability without promising arbitrary-event boundaries. Please update the JSON description, regenerate README, and qualify the equivalent “any step” wording in the PR body.

Actual Node 24.19.0 checks on this head: 120 projects / 10 categories validated, two generator passes were byte-identical to the committed README, the reviewed diff passed whitespace checking, and the initial full link pass was 120 ok / 0 failures / 0 warnings. A later full check of current upstream main found 118 ok / 1 failure / 0 warnings due to HTTP 404 on existing Fzkuji/OpenProgram; that separate catalog availability issue also prevents using the earlier green result as a current all-links-clear result. OrcaReplay's URL passed, its Apache-2.0 license was present, and the normal weekly star snapshot drift is not a requested change.

After updating the description:

node scripts/generate-readme.js
node scripts/validate.js
git diff --check
node scripts/check-links.js

Both Actions runs on this PR are currently action_required with 0 jobs (validate, links); Cursor Bugbot success is not a replacement for these checks. Maintainer workflow approval and successful execution are still required. No merge has been performed.

[cursor]

Bugbot is paused — on-demand spend limit reached

Bugbot uses usage-based billing for this team and has hit its on-demand spend limit.

A team admin can raise the spend limit in the Cursor dashboard, or wait for the next billing cycle to continue.

[xizhuomengcontin]

Thank you for reading the implementation rather than the description — the finding is correct and I have changed it.

data/projects.json description now reads:

Records a coding agent's exchange with its model provider and replays the run without calling the provider; replay can fork from available checkpoints onto a different model. Recorded tool calls still execute on replay.

Two things changed, not one:

1. "fork from any step" → "fork from available checkpoints." Your reading of deriveCheckpoints and the fork command matches what the code does: a checkpoint needs both a complete conversation prefix and a filesystem snapshot, --fork refuses a run with none, and a non-checkpoint --from is snapped back with fork.snapped. "Any step" promised a boundary the tool does not offer.
2. Dropped "offline" from the same sentence. Replay withholds the model provider; it does not take the run off the network, because recorded tool calls still execute. That was a second overclaim in the same line and it should not have survived into the entry either.

I also qualified the equivalent wording in the PR body, including the --from example, and removed the hardcoded star count from the body so it does not go stale.

Checks on this head, Node 22.23.2:

node scripts/generate-readme.js   -> regenerated
node scripts/validate.js          -> Validated 121 projects across 10 categories / Validation passed.
git diff --check                  -> clean
node scripts/check-links.js       -> Results: 121 ok, 0 failures, 0 warnings

Fzkuji/OpenProgram resolves again, so the full link pass is green rather than 118 ok / 1 failure.

The branch is rebuilt on current main — it had drifted into a conflict since September, and the PR is now a clean +13 −0 in the JSON plus the one generated README line. The Actions runs still need maintainer approval to execute.

Disclosure: I maintain OrcaReplay, so this is a self-submission.

[autogame-17]

Rechecked the revised head 2eca16c61d6965ba13b028305192afb2e493d126 on 2026-10-12. The previously requested description fix is addressed. data/projects.json:1569 and the generated README.md:157 now say “available checkpoints” and explicitly disclose that recorded tool calls execute during replay. This matches current upstream a9178ca26af95cfd3231b8eb2da4ae11dbbd0a84: deriveCheckpoints and the fork path enforce the checkpoint boundary.

Actual local verification using the current trusted main scripts on this head's data and README, with Node 24:

• Validation: 121 projects / 10 categories; passed.
• Two generator passes: byte-identical to the committed README.
• Whitespace check: passed.
• Merge simulation against current main 596c20ba06738d5d7fcb10afd9934e507b4b81bd: clean.
• The canonical repository is accessible, active, Apache-2.0 licensed, and neither archived nor disabled (285 stars at this check).

The remaining blocker is CI execution, not the old wording: Validate curated data and Check Links are both action_required, with 0 jobs. A maintainer must review the workflow execution boundary and authorize those runs through the normal GitHub process; both checks must actually succeed before approval. The PR's locally reported link pass and Bugbot's billing-paused state are not replacements for that evidence.

No further description edit is requested by this follow-up. Keeping the PR open/blocked; no approval, merge, or workflow execution was issued by this maintenance run.

[autogame-17]

Verified the refreshed catalog-only diff against current main. The previously corrected available-checkpoint boundary is retained without additional wording changes; the title now accurately describes provider-exchange replay. Generated README, validation, and link checks passed locally and in the two current GitHub Actions jobs.