Conversation
Owner decision after the seventh update (614796841) went out at thirteen bullets: "don't feel obliged to make the list long, it's more important to keep it interesting and mention significant results." Every bullet there had passed all three gates — which is the trap the docs did not name: the gates decide what is allowed, not what is worth reading. SKILL.md §What counts now states that eligibility is not entitlement and that padding is forbidden; §6 adds that a short message is a normal outcome after a large window. HANDOVER.md leads the house-style section with the rule and records 614796841 as the calibration point in the wrong direction. Also refreshes the state table: watermark 54d9297/1770, the seventh message and yesterday's in-place lSeries edit, and the cron id caveat. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019SoB4BZFGPhMAJWfvnLXCo
Three lessons from the 2026-08-07/08 runs, on top of the curation rule. - Watermark protocol gains rule 5: copy the full SHA from `git rev-parse`, never pad an abbreviated one. On 2026-08-08 the state DM went out with a plausible 40-character string naming no object, which would have made the next window unresolvable; the remedy is a superseding DM, and `git cat-file -t` is the cheap check. - Significance gate: the gate judges only the summary it is given, so a summary written in the language of the roadmap milestone rather than of the file's own Main results can return ANNOUNCE for something unproved. Worked instance: an adic-spaces batch described as "establishing that the valuation spectrum is a spectral space" earned that recommendation, while no SpectralSpace conclusion exists in the tree — the patch criterion and pro-constructible calculus had landed, the endpoint had not. Same Bochner trap, arriving through the gate. - Scheduling: say plainly that a session cron's firing time is approximate. It queues into the session's own loop, fires only when that loop is idle, and carries up to 15 minutes of jitter. Observed 18:42 and 16:33 against a 16:03 slot. Don't promise a wall-clock minute; that is an argument for Actions, not for tuning the cron. Ops gotchas also gain the Python-TLS failure (python.org build lost its CA file, so `zulip.py check` fails while curl works — the run path is curl-only) and a note that a freshness abort is a successful outcome. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019SoB4BZFGPhMAJWfvnLXCo
Chris, 2026-08-09: "there is no rule to cut it to 5. the rule is that there doesn't need to be a minimum or maximum, but you just need to make them be interesting." The 2026-08-05 curation rule was written after a thirteen-bullet message and, in being applied, hardened into a de-facto cap of five. On 2026-08-09 that cap cut a legitimately interesting item — the Huber rings, Tate rings, Huber pairs and strong-noetherianness block that is Layer 0 of the adic-spaces roadmap, which would have paired with that same message's lead bullet on the spectrality of Spv. Both files now say plainly that there is no minimum and no maximum, that any "a handful" phrasing describes past windows rather than setting a quota, and that cutting a bullet you would have been glad to read means the count is driving. HANDOVER records both calibration points: 614796841 padded to thirteen, 615474670 trimmed to five. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019SoB4BZFGPhMAJWfvnLXCo
Chris, 2026-08-10: "you cant rely on pr title, dont be lazy, this is
meant to be a slow methodical and careful check. you need to look at
the contents of each PR and from that make your list."
The old §2 said to triage by title prefix and only open the feat: ones
that sounded like announcements. Applied as written, that produced the
worst miss so far: fifteen feat(EllipticCurve) PRs across two windows,
none opened, zero elliptic-curve results in 105 announced bullets —
among the casualties Silverman III.10 on Aut(E) for j outside {0,1728}
(#2248), the quadratic twist and its invariant theory (#2254), and
quadratic Galois descent (#2268).
§2 now requires fetching and reading every body in the window, all
prefixes, and deciding from what the body says was proved or defined.
It also states that the window bounds discovery, not eligibility: the
dedupe keys on cited PR numbers, so a result found late is announced
late, not never. HANDOVER records the operator lesson with the quote.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019SoB4BZFGPhMAJWfvnLXCo
…lity Chris, 2026-08-10: make sure ChatGPT 5.6 Sol via the MCP gives a second opinion on how interesting/noteworthy a result is. The gate section read as a binary ANNOUNCE/SKIP filter; in practice its per-item reasons and ranking are what curation runs on. The prompt template now asks for a noteworthiness ranking of the strongest candidates, and a new paragraph states that the ranking decides which result leads and that curation is made with the gate's reading in hand, every pass, previews included. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019SoB4BZFGPhMAJWfvnLXCo
A macOS LaunchAgent (com.tauceti.voyager, 16:03 local) now fires each daily run as a fresh headless Claude session. Session crons carried two production failure modes: they die with their session (the 2026-08-11 credits incident forced a takeover mid-week), and macOS App Nap delayed every idle-session firing by 20-40 minutes past the slot. skills/voyager/launchd/ carries the installed plist, wrapper, headless prompt (the daily prompt minus the now-unneeded cron-upkeep step, with ask-Chris rerouted to the run log), the read-only smoke test, and a README with the operating commands. HANDOVER.md's scheduling row now describes the launchd arrangement. Verified on the operator machine: plutil lint, a full headless smoke run (gh keychain, Zulip read, MCP availability, docgen fetch — all pass), and launchctl registration with the 16:03 calendar trigger. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The launchd runner now walks .claude3 -> .claude2 -> .claude4 -> default and stops at the first account that completes the run, so an exhausted usage limit on the primary no longer costs a day's update. Exit status cannot drive this: `claude -p` exits 0 even when it never ran a turn. Both failure modes were observed live while building it — the default account was at its weekly limit and .claude5's organisation has Claude Code subscription access disabled outright, and each printed one line and exited cleanly. The chain therefore keys off two positive signals: a VOYAGER_RESULT sentinel the prompt now requires on every terminal outcome, and freshness.py, a read-only check for a Voyager post or state DM inside the 2h window that also runs before each attempt so a run which posted but died before printing its sentinel is not repeated. Retrying is safe by construction — the skill's freshness abort makes a second attempt exit quietly rather than double-post. Verified end to end: with the chain overridden to .claude5 then .claude2, the org-disabled account was detected in three seconds and .claude2 ran the full environment smoke test and reported success. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
On 2026-08-15 no update went out. Both codex accounts were unavailable -- ~/.codex on quota until 2026-08-20, ~/.codex2 on quota until 2026-09-14 and additionally refusing gpt-5.6-sol, gpt-5.4 and gpt-5.3-codex with HTTP 400 -- and the docs made the ChatGPT significance gate a precondition for posting. Two runs read all 138 PR bodies in the window, assembled a slate, and published nothing. Holding is the worse failure. A window that plainly contains notable results and produces silence is indistinguishable to readers from a quiet week, and the watermark protocol already makes a *missed* day harmless. So: * SKILL.md §4: the gate is best-effort. Probe it before reading the window, so an outage costs seconds rather than a full read. If it answers, its ranking drives bullet order as before. If it does not, run without it -- select and order on judgement, post as normal, log the reason. Never substitute another model, and never mention the gate in the Zulip message; that belongs in the log. * SKILL.md Prerequisites: the "stop and report if credentials are missing" rule governs Zulip and gh. It never governed the chatgpt-math MCP, but it could be read that way, which is part of how the gate came to look mandatory. Also tracks the scheduler, which until now existed only in ~/.claude3/voyager/ on one machine, with no review trail and no backup: prompt.txt (whose standing rules override this file), run.sh, freshness.py, status.py, smoke.txt. freshness.py and status.py are included because run.sh calls the first and voyager-status is the second -- a snapshot without them would not restore a working system. The deployed copy stays at ~/.claude3/voyager/; scheduler/README.md records how to sync and why the fallback logic reads the way it does. run.sh carries one fix in this snapshot: account health is decided by the VOYAGER_RESULT sentinel, never by grepping the transcript for limit phrases. Two healthy accounts had described *codex's* quota in their own reports, matched that grep, and were logged as rate-limited Claude accounts -- so the chain burned every account on a blocker no switch could fix, and voyager-status showed four dead accounts when two were fine. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Every channel message's bold header now ends with the italic suffix *(AI-generated summary)*, per Chris's request of 2026-08-19. Applies to the full update and the quiet check-in alike, and survives codepoint-cap trims. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The chain is now models x accounts: every account is probed for claude-fable-5 (one trivial completion — the only headless test of 'has credits for this model right now') before any attempt falls back to claude-opus-5. Owner request, 2026-08-19. Also consolidates the run.sh mirror: scheduler/run.sh had drifted ahead of launchd/run.sh (the 08-15 sentinel-first fix landed only there); the single canonical mirror is now the documented launchd/ bundle, synced to the live file, and the stray copy is removed. launchd/prompt.txt is synced too (gate-best-effort clause and the AI-authorship header rule). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Voyager skill-doc corrections from the 2026-08-05..10 production runs
Five commits of prose updates to
skills/voyager/SKILL.mdandskills/voyager/HANDOVER.md— no other files touched. They codify the owner corrections and production lessons accumulated while operating the daily update, so the next operator inherits them from the docs rather than from session memory.The three behavioural corrections (owner instructions)
chore:/refactor:/fix:included); titles may order the reading, never exclude a PR. Calibration: fifteenfeat(EllipticCurve)PRs missed by title-tone triage across two windows (TauCeti#2248, #2254, #2268 among the casualties). Also states that the window bounds discovery, not eligibility — late-found misses fold into the next run.Also carried
git rev-parse, verify withgit cat-file -t(the 2026-08-08 incident).## Main results, or the gate approves an unproved endpoint (the Spv-is-spectral near-miss).certifigotcha, and "a freshness abort is a success".Machine-readable records of the same lessons: #28 (
data/community_learnings/).🤖 Generated with Claude Code