nyldn/claude-octopus v11.9.5

v11.9.5 — Keep long sessions spawning seats on Linux, stop argument substitution rewriting commands, and let research cite workspace files

on GitHub

Fixed

• Seat spawns no longer fail once a session's seat ledger passes 128 KiB. The
  run manifest writer handed the whole seat projection and the event list to
  jq as --argjson values, one argv string each, which Linux caps at
  MAX_ARG_STRLEN (128 KiB). A session's ledger
  (runs/<session-id>/seats.jsonl) keeps every run's records, so after a few
  reviews the projection outgrew the cap, jq failed with E2BIG, every
  transition rolled back with "Unable to persist planned execution contract",
  and each run ended with "ALL Round 1 providers failed" until a new session
  started (#1111). Both ledgers now reach jq through --slurpfile, so
  manifest publication no longer depends on the ledger's size; the manifest
  content is unchanged.

• Provider-level reasoning effort no longer collides with the codex
  reasoning model slot in providers.json. The v3.0 config Octopus writes
  stores a model at providers.codex.reasoning, used by the codex-reasoning
  agent and the security: codex:reasoning route, while role-based execution
  profiles (#616) read a {default, policy} effort object from the same key.
  On a v3.0 config that lookup failed silently; storing the object instead made
  codex-reasoning and codex dispatch in the security phase fail with
  "Invalid configured model". Provider-level effort now lives in
  providers.<provider>.reasoning_effort and
  providers.<provider>.reasoning_policy, and reasoning stays the model
  slot. An effort object already stored at reasoning is still honored, and
  the capability-map lookup now skips non-string values, so the object is
  never taken for a model name. (#1114)

• /octo:auto requests for parallel work, such as "decompose the auth refactor
  into parallel work packages", now hand off to /octo:parallel, the same as a
  confirmed --workflow parallel choice. Since 11.4.0 the automatic router
  passed the request text to the tasks-file runner behind
  orchestrate.sh parallel, so these requests stopped with
  Tasks file not found: <request> and exit 1. (#1116)

• Commands and skills invoked with arguments no longer have their shell and awk
  code rewritten by Claude Code's $N argument substitution. /octo:embrace lets do it all rendered the provider banner's command -v "$1" as
  command -v "do", a shell keyword, so every CLI was reported available, and
  flow-parallel's launch script resolved dirname "$0" to the first argument.
  Shell parameters now use ${N}, awk fields use $(N), and literal amounts
  such as cost estimates use Claude Code's \$ escape. A unit suite rejects any
  unescaped $N in command and skill bodies, the Cursor copies, and the shipped
  skills. (#1101)

• /octo:embrace debate gates no longer pick their context from the shared
  ~/.claude-octopus/results/ root. orchestrate.sh writes phase documents to
  the session's results directory, so ls -t over the root found no current
  document and returned the newest match left by an earlier session. The
  gates now pass no context file and let orchestrate.sh read its own session
  directory, and the command uses the document paths the phases log. (#1101)

• Research about the local codebase can now pass evidence verification. A
  claim may cite a workspace file as path:LINE, path:START-END or
  path:A,B. The citation counts when the file exists inside the workspace
  root recorded in the run manifest and every cited line exists, and quotes and
  numbers in the claim must then appear in that file. Before, every claim
  without a web [source:S###] ID failed, so a probe that researched the
  repository could never publish its synthesis. (#1102)

• URLs harvested from provider output no longer keep a JSON-escaped \n, \r
  or \t, or a trailing backslash. (#1102)

• Probe synthesis reads as much of each provider artifact as the synthesizer's
  context budget allows (for example OCTOPUS_CLAUDE_CONTEXT_BUDGET). The
  fixed limits of 24000 bytes per artifact and 120000 in total cut Codex
  answers off after their prompt echo. OCTOPUS_PROBE_SYNTHESIS_FILE_CHARS and
  OCTOPUS_PROBE_SYNTHESIS_CONTEXT_CHARS still pin either limit. (#1102)

• orchestrate.sh --dry-run fan-out no longer reports every agent as a failed
  spawn. A dry-run spawn prints the command it would run but no provider PID,
  and fan-out still waited for one, so it logged a PID error and a failed spawn
  for each agent and then an empty "All agents spawned" list. Dry-run fan-out
  now prints one preview per agent and stops there; an agent whose command
  cannot be rendered is still reported. (#1113)

Thanks to @pdrgds (#1101, #1102, #1113, #1114, #1116) and @vkchan (#1112) for these fixes.

Full Changelog: v11.9.4...v11.9.5