Every claude-code run ends in error_max_turns, and `partial` stops the fallback chain before the third tier

[rssrn]

Summary

Every claude-code run ever recorded has ended partial with error_max_turns. Four for
four. Because partial counts as a result rather than a failure, it terminates the
fallback chain — the third tier is never reached — and the run posts a comment that says
only that the agent failed.

Combined with #64 (codex has not authenticated since July), this means the proactive lane
has produced nothing usable since early July, at roughly $1 per attempt.

Evidence

agent	model	outcome	failure_reason	n	last
claude-code	claude-opus-4-8	partial	error_max_turns	4	2026-08-28T06:25:05Z

There is no success row for claude-code at any point.

The 2026-08-28 run (b5a7c4eb, issue #63):

06:25:15  invoking agent claude-code (slug=claude-code:claude-opus-4-8)
06:26:56  WARNING claude error subtype: error_max_turns
06:27:00  labelled issue #63: +ai-failed +ai-contributed
06:27:00  run complete: outcome=partial dur=116s turns=16 cost=$1.0714 tokens=471,909

turns_used=16 against max_turns=15, actions_taken=[], wip_branch_url empty. 101
seconds of agent time, $1.07 and 472k tokens spent to produce the comment "investigated
this suggestion but reported failure. Reason: error_max_turns".

Two distinct defects

1. max_turns = 15 is too low for the proactive persona

[defaults] max_turns = 15 in labro.toml, with no override on any
proactive-improvement task source. A perspective audit asks the agent to explore an
unfamiliar repo, form a thesis, and then act on it. Fifteen turns is plausible for a
labelled-issue task with a known target; it is not enough to read a codebase from cold.
The 100% failure rate across four runs and two repos suggests this is systematic, not bad
luck.

Worth deciding whether the fix is a higher max_turns on the proactive sources
specifically, or a tighter prompt that narrows what the agent has to read before it can act.

2. partial terminates the fallback chain

The proactive model list is:

model = ["codex", "claude-code:claude-opus-4-8", "opencode:opencode/deepseek-v4-flash-free"]

Tier 3 has never run. Codex 401s (a hard failure, so the chain advances); claude-code
returns partial (a result, so the chain stops). The cheap free-tier fallback that exists
precisely for this case is unreachable in practice.

Whether error_max_turns should advance the chain is a genuine design question — a
partial result carries real work and re-running from scratch discards it. But the current
behaviour means a permanently-degraded tier-2 agent silently consumes every proactive run
and the configured safety net never engages. At minimum error_max_turns with
actions_taken=[] is indistinguishable from a failure and should be treated as one.

Related

• Codex auth is structurally broken: single-use refresh token stored as a snapshot secret, no successful run since 2026-07-03 #64 — codex auth, the reason tier 2 is being reached at all
• skipped masks crashes and total source failures — a broken run is indistinguishable from an idle one #60 — outcome reporting hides broken runs; this is the same class of masking one layer up
• Remove WIP branch preservation and resume — 3 branches created in 1,999 runs, 0 resumes ever #62 — a partial run with an empty wip_branch_url is more evidence that WIP
  preservation is not doing the job it exists for