Historical record
This page records what happened on a date, with its observed failures and limits. It is not a requirement for the current implementation. Current behaviour is in Status.
Returning to a stopped P3 analysis
2026-11-02. This pass tests whether a person can find a real conversation after closing its tab and continue after Claude Code reaches its per-message turn limit. It is not an unattended-run or scientific-reproduction acceptance.
Before the change
An isolated project copied the P3 README and the public GSE232853 processed matrix (16,560 genes and 578 AOIs). It used the same original question, Claude Code through claude-agent-acp 0.97.0, the configured Haiku 5.5 model, Medium effort and Rho's 64-turn limit. The service used its own state directory, native Agent and Notebook, and no Core Host. A small harness chose the Agent's offered one-time permissions; those choices are not unattended execution.
The baseline used the existing Rho 0.5.2 build while the checkout was at 1ae9d07f; the build was not independently tied to that source commit. The continuation used the local application changes in this pass and the selected Agent plugin 0.4.25 at Rho-plugins 90e68cfada22fe939ba5e47180823a8372370771. Node was 22.22.3 on macOS arm64. This is local composition evidence, not an installed-package acceptance.
In a real Chrome tab the conversation row expanded details but did not open the conversation. The page address did not identify the project or conversation. The Agent reached Internal error: Reached maximum number of turns (64) after 104 recorded steps and 19 distinct one-time choices. Its original turn remained failed with its steps and partial files. The new run did not reproduce an identical sequence of calls, but reproduced the original reported stopping error.
Change and result
The row now opens the retained conversation; its fragment identifies the project and conversation, not a credential. Opening the address reads the record and starts no Agent. The first query-parameter implementation failed because the service refuses that browser entry; the fragment passed in a separate Chrome context and in another tab observing the real P3 run. Pairing codes still leave the address when redeemed.
The panel, list and Overview name a confirmed Agent turn limit as such. An Agent ending its reply is no longer presented in the panel as proof that the analysis finished. A person can explicitly continue the latest confirmed limit stop. Rho retains the original turn and creates a new message with its own request identity, linked to that turn and supplied with the original task. The action checks the latest retained state and cannot continue an unknown or stale turn. Duplicate requests with the same identity return the original message. A session-not-found error after dispatch is unknown and is not automatically retried.
The owned service was closed and restarted using the same isolated project and state. A fresh browser tab reopened the stopped turn without sending a message; the person-equivalent action selected Continue unfinished work. The retained native ACP session loaded with the same session identity. A new t2 linked to t1 finished with end_turn after 116 seconds, ten retained steps and four further one-time permission choices. The script, marker and differential-expression tables, Hallmark table and dotplot, and findings file were present afterward. The old t1 and its 104 steps remained. No scientific result was independently verified in this pass.
The service restart did not preserve the R Notebook's in-memory objects. Despite the original request to work in the shared Notebook, the Agent chose to run the saved script with Rscript; its final answer disclosed that the final files came from the script. The four additional choices were answered by the harness. These facts limit the conclusion: the observed path restored a conversation and continued its task, but did not demonstrate uninterrupted Notebook execution, correct scientific conclusions, or unattended completion. The Agent's claims about the analysis remain its claims.
The initial stop and completion screenshots and local run records are under work/continuity-repair/; they contain the actual test project and are not part of the source release. The first stop screenshot also exposed an Overview label that still said Failed; the Overview was changed to use the retained limit fact.
Checks
node --test test/acp-conversations.test.mjs test/conversations.test.mjs test/conversation-presentation.test.mjs test/protocol.test.mjs: 23 passed. The ACP peer is deterministic; it checks identity, restart, unknown outcomes and no duplicate dispatch, not real model behavior.node scripts/agent-page-acceptance.mjs: real Chrome with a controlled ACP peer; the run checks row opening, a fresh browser context and the limit presentation as well as the existing Agent-page paths.npm run check:contracts: checks the generated public contracts and TypeScript.
The one-click continuation path is intentionally limited to explicit turn-limit reports. Wall-time cancellation, a lost reply, a cancelled native tool and the Agent's assertion that background work will wake it are not evidence that an analysis is complete or safe to replay. Persistent authorization, resource-path discovery and real background wakeup remain separate work.
Defaults chosen afterwards
The user then chose Claude Code's Auto permission mode as the default and no turn limit. Rho now sets no step cap (maxAgentTurns) and no message wall time (turnWallMs) unless a --limits file names them; output, line and record limits remain, and the person can stop a message. Agent 0.4.26 (Rho-plugins 2a9cd421) passes Claude Code a step cap only when one is set, and applies preferred settings to a new session. Rho prefers mode: auto over Manual (default); a mode chosen in Rho or named in the person's Claude settings stays, and a loaded session keeps its own, so this pass's P3 conversation would continue in Manual. The pinned claude-agent-acp 0.97.0 offers Auto and falls back to Accept edits, with a warning, when the model does not support it. Codex's adapter already opens in its auto-review mode.
A stand-in ACP peer checks the result: a new session receives no maxTurns and is switched from Manual to Auto; a chosen mode, a mode from the Agent's own settings, an unoffered value and a loaded session are left alone; a configured cap of 64 is still passed.
Running P3 again with the defaults
The user chose to let the harness answer any remaining permission request with its one-time allow and started it themselves, since the permission check here would not. Each run copied the README and matrix into a new /tmp project, started an isolated service and sent the original question with only the original Medium effort; nothing else was configured. The session reported auto on fusecode-lab/claude-haiku-5-5[1m] through claude-agent-acp 0.97.0 both times.
| Run | Bridge | Steps | Time | Permission requests | Outputs required by the README |
|---|---|---|---|---|---|
| Before (above) | 30 s | 104, then 10 | stopped at 64 turns, then 116 s | 19, then 4 | after the continuation |
| 1 | 30 s | 62 | 4.2 min, end_turn | 0 | 1 of 6 when the reply ended |
| 2 | waitMs + 30 s | 34 | 12.9 min, end_turn | 0 | all 6, in the shared R session |
Run 1 ended without finishing. The limma cell was sent with waitMs 120000, as Rho's instructions suggest, but the Agent's MCP bridge ended every call to the service at 30 s, so the call came back as an unknown outcome while the cell went on running. The Agent interrupted it, started nohup Rscript scripts/analysis.R in the background (Claude Code refused sleep and pointed to Monitor), and ended its reply saying it was waiting on a Monitor. Nothing resumed it. The orphaned script wrote the DE table, Hallmark ORA and dotplot about five minutes later; notes/findings.md was never written.
Agent 0.4.27 (Rho-plugins 1eb1d3e9) gives a call the waitMs it asked for plus 30 s. In run 2 a cell waited 259 s and returned. Another, sent with waitMs 600000, still came back unknown after exactly 300 s: Node's fetch stops waiting for a reply's headers then, and the service answers a waited cell only when it ends. The Agent read the session, found the cell running and sent the same change identity again, which returned that cell rather than a second one, and went on. The bridge now uses node:http; a 310 s cell sent through it with waitMs 400000 returned done after 310 s. Whether Claude Code's own MCP call limit allows such a call was not observed.
Run 2 stayed in the Notebook throughout, wrote the six outputs and a paired-only alternative, and checked its own claim that the alternative was "essentially unchanged" before keeping numbers (1,043 shared DEGs, logFC correlation 0.995). For ten minutes after the reply the service stayed up and nothing changed. Its counts equal the earlier literature run's: 1,111 DEGs (833 up in CD68, 278 down) and 13 of 50 Hallmark sets, led by allograft rejection (BH 2.8e-7). That is agreement between runs on the same data, not a check against the paper, which the README gives no numbers for. One run each is not a rate; the harness, not a person, would have answered any request.