Codex host guidance and honest connection failures; halve the configuration docs

[tobert]

A Codex MCP call can reach a provider while the same kaibo CLI call fails inside Codex's command sandbox. kaibo used to report that connection failure as a provider rejection. This PR classifies it as a connection failure, points the caller at the checks to run, adds a short host-agents section to the configuration reference, and halves the configuration docs.

What changed

• Connection failure class (src/server/render.rs). rig flattens reqwest's chain to HttpError … error sending request; that now reads as "no response from the provider, not a rejection", names the endpoint/DNS/proxy/TLS/network checks, and points at Host agents in kaibo://config/guide. Timeouts and resets stay TransientProvider (the transient markers are checked first, and the test now asserts the classification, not just the wording).
• Every failure message tells the calling agent to ask the user what to do — retry, use another cast, or continue without the consultation. Main's messages told the agent to "proceed without the consultation", and the first draft here swapped that for "within the user's authorized scope". Amy: "I don't think we should tell an agent to proceed without consultation, it should ask the user what to do." A new test takes one body per failure kind, asserts its classification first, and requires "Ask the user whether" with no "proceed without"; it failed before the messages changed. The kaibo-side message no longer says "please report it". The transient message states TRANSIENT_RETRIES and that a timeout or reset is not retried, and the README failure paragraph no longer claims kaibo never retries.
• kaibo config-guide prints the embedded reference without loading config or telemetry, so it works when the config is broken. The test runs the binary with a cleared environment and a malformed config and compares stdout byte for byte.
• Host agents, not a separate Codex guide. Amy: "I feel like codex.md should merge into the config guide, which should have a short section for per-agent quirks." So docs/codex.md, the kaibo://config/codex resource, and the config-guide codex topic are gone; ## Host agents / ### Codex at the end of docs/config.md keeps what changes setup: the launch snippet with an approval default, separate MCP and shell network access, forwarding api_key_env variables, the 60 s tool timeout and the async tools, store write access for shell-run kaibo. Copies of Codex's own auto-review, elicitation, persistent-approval, and permissions-profile documentation were cut: they drift with Codex, the approval UI was unverified, and kaibo sends no elicitation requests.
• The resident instructions lead says "codebase review and second opinions from another model family" and discloses that hosted casts send source to providers. It no longer names a Codex resource, which every host would pay for each session.
• Configuration reference and template: 23,669 → 11,223 words, including the new Host agents section. A Sonnet audit compared every config key, KAIBO_* variable, and flag against the structs and found none lost. Restored from the cut: a blocked disable_builtins call exits 126 and kaibo serve refuses an unknown name, a write past scratch_limit_bytes is refused, a scratch-directory allow_paths example, and the configure prompt's OpenRouter public-catalog query. Two audit items were stale rather than lost and stay out: the "no default root" case (the inferred cwd is additive now, resolver.rs) and a bare $TMPDIR example (an unset variable fails loading).
• Doc errors fixed while there: --config beats $KAIBO_CONFIG (the old block had it backwards, inherited from main); reasoning.effort on the Responses wire is sent only for gpt-5* ids; telemetry has no CLI flags.
• Merged main (A slow model call shows in job_get and warns the caller #196): slow_chat_secs is one line in the short template.

Not in this PR

• classify_failure has no marker for a bare 500/502 body, so a call completion_retry just retried four times reads "retrying is unlikely to help". Pre-existing; its own PR with a failing-first test.
• MCP tool annotations, a planned follow-up PR (Amy: "we can do a followup PR with tool annotations, that probably will help modern claude code too"). kaibo sets none today, so Codex's approval_mode = "writes" prompts for every kaibo tool; readOnlyHint would let it auto-approve paid calls, so that PR weighs it against openWorldHint.
• A Claude Code subsection under Host agents.
• [sandbox] disable_builtins names are validated only in kaibo serve (main.rs:201); a CLI run with a misspelled name silently leaves the builtin enabled, although sandbox.rs says it was validated loudly. The template now says kaibo serve; the CLI gap is its own fix.

Validation

• Offline suite: 1,377 passed, 0 failed, 23 ignored (live provider tests); cargo clippy --all-targets -D warnings clean.
• The new ## Host agents heading guard fails when the heading is renamed (negative control run, then restored byte for byte).
• every_failure_asks_the_user_what_to_do failed on the old messages (exit 101) before they changed.
• Codex CLI 0.154.0 probes (first commit): MCP connectivity versus the sandboxed CLI, and a command-domain policy that reached an allowed provider host and refused a denied one.

Review

• Codex (GPT-6 Astra) wrote the first three commits; DeepSeek Flash reviewed them through kaibo and caught the connection classifier absorbing timeouts.
• Claude Opus 5 reviewed the PR for what helps Codex use kaibo, then made the fold and the fixes above. DeepSeek Flash (kaibo deepseek cast) verified the doc corrections and the Codex guide's kaibo-side claims against the code, and reviewed the final revision.

🤖 Generated with Claude Code