Repository navigation
Attempts start with the target's research memory - #191
Conversation
Reports were archived on the research-log branch but nothing read them back: agents re-tried yesterday's negatives this morning. A fresh attempt now fetches the branch (fail-soft: no branch, no network — empty memory, never a dead attempt), the brief inlines the newest reports with a standing instruction not to repeat settled experiments, the full texts land in the channel archive, and `syscall reports` lists them one line each or prints full reports, several per call. A second cluster's kernel reads the same branch, so this is also the cross-cluster findings sync; the dev plan for cross-cluster deployment is in the roadmap. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A failed attempt is not a dead idea: one point in a parameter space, a walltime kill, or an infrastructure failure closes nothing. The memory section now says what a report settles and what it leaves open. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
There was a problem hiding this comment.
Round 1 — reviewed head 27164b60 — reviewer summarizer:hermes/gpt-5.6-terra over coverage+credentials+deployment+general+lifecycle+prose.
terra
Advisory findings from autoresearch — the code owner decides. Reply to disagree; the autoresearch:no-review label opts this PR out.
Verdict: 1 blocking, 3 advisory.
4 findings attached to the lines below.
Blocking: syscall installation deletes the fetched report archive before an author session can use it (coverage, credentials, deployment, general, and lifecycle agree). Suggestions: make the archive test reproduce the production ordering; ensure the advertised reports command exists for every advertised path; and bound each remotely sourced report. Rejected findings: none; the five identical archive-loss opinions were merged as one finding rather than dropped.
There was a problem hiding this comment.
Round 1 — reviewed head 27164b60 — reviewer hermes/gpt-5.6-terra.
terra
Advisory findings from autoresearch — the code owner decides. Reply to disagree; the autoresearch:no-review label opts this PR out.
Verdict: 1 blocking, 0 advisory.
1 finding attached to the lines below.
The full report archive is removed during setup, so the advertised report reader cannot access it.
The syscall installer recreates the channel dir it owns, so the archive is written after it (it was deleted on every author-syscall attempt); the test now exercises that order. The brief advertises the reports command only when the tool is installed; reports still inline without it. Each fetched report is capped at 100k characters — the branch's content is remote-controlled. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
There was a problem hiding this comment.
Round 2 — reviewed head f535e1b4 — reviewer hermes/gpt-5.6-terra.
terra
Advisory findings from autoresearch — the code owner decides. Reply to disagree; the autoresearch:no-review label opts this PR out.
Verdict: nothing blocking — 1 advisory note.
Advisory (non-blocking):
- Nested report paths can overwrite another archived report. The fetch flattens every remote reports/*.md path to its basename, so reports/root.md and reports/nested/root.md both become root.md and the later write replaces one full report in the archive. (
src/autoresearch/attempt.py:158; high confidence)
Nested report paths can overwrite an archived report.
…n onto another report Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
There was a problem hiding this comment.
Round 3 — reviewed head c56916f0 — reviewer hermes/gpt-5.6-terra.
terra
Advisory findings from autoresearch — the code owner decides. Reply to disagree; the autoresearch:no-review label opts this PR out.
Verdict: nothing blocking — 1 advisory note.
1 finding attached to the lines below.
A new brief-only input is silently ignored on resume.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
There was a problem hiding this comment.
Round 4 — reviewed head 048c6a10 — reviewer hermes/gpt-5.6-terra.
terra
Advisory findings from autoresearch — the code owner decides. Reply to disagree; the autoresearch:no-review label opts this PR out.
Verdict: 1 blocking, 0 advisory.
1 finding attached to the lines below.
The archive size cap is applied after the remote blob is already read.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
There was a problem hiding this comment.
Round 5 — reviewed head f5f93c31 — reviewer hermes/gpt-5.6-terra.
terra
Advisory findings from autoresearch — the code owner decides. Reply to disagree; the autoresearch:no-review label opts this PR out.
Verdict: 1 blocking, 0 advisory.
1 finding attached to the lines below.
The new reports pointer is workspace-relative, so it fails under Hermes.
Summary
Reports were archived on the
research-logbranch but nothing read them back — agents re-tried yesterday's Muon negative this morning. Now:reports/(newest first). Fail-soft: no branch or no network is an empty memory, never a dead attempt..autoresearch/reports/(bounded at 30; names sanitized — branch content is remote-controlled).syscall reports= summary list, one line per report (name + Outcome);syscall reports <name> <name> …= full reports, several in one call.docs/roadmap.md.Test plan
🤖 Generated with Claude Code