Troubleshooting

A review times out

Symptom: “Provider did not answer within 600 s — raise ‘Request timeout’ in settings if your model needs longer.”

Slow local models are the usual cause. A 7B model on CPU, or any model with thinking enabled, can legitimately take many minutes on a long note.

1. Raise the timeout. Settings → AI Editor → Behavior → Runs → Request timeout (seconds). Default 600, maximum 3600.
2. Turn thinking off while you diagnose. Backends → your backend → Thinking. A model reasoning silently for minutes is indistinguishable from a hang, which is why it defaults to off.
3. Review a selection instead of the whole note to see whether size is the problem.
4. Lower the context budget or turn off Include linked notes on the editors involved. Fewer characters, faster answer.
5. Reduce concurrency. Max concurrent requests at 3 means three models loading at once on the same machine; drop it to 1 for a local backend.

A CLI backend ignores that setting. It carries its own Timeout on the backend itself (default 300 s). Raising the Behavior tab’s request timeout does nothing for Claude Code or Codex.

Test connection uses its own 60-second bound, so a local model still loading its weights can fail the test and work fine for a real review. The message says so when that is what happened.

Requests fail immediately (network error, or “Failed to fetch”)

Symptom: “Network request failed: … Check the endpoint URL and that the server is reachable; if it is a self-hosted endpoint, it may be blocking browser requests (CORS)…”

Requests are made from Obsidian’s renderer, which is a browser context. The browser deliberately hides which of DNS failure, connection refusal and CORS rejection happened, so the message names all three.

• Ollama — it must allow Obsidian’s origin. Start it with OLLAMA_ORIGINS=app://obsidian.md, or add that to its service environment permanently.
• LM Studio, llama.cpp, vLLM, a corporate gateway — same class of problem. Enable CORS for app://obsidian.md, or put a permissive reverse proxy in front.
• Check the base URL — a trailing path, a wrong port, https where the server speaks http. Trailing slashes are handled for you.
• Check the server is up at all: curl the same URL from a terminal. If curl works and Obsidian does not, it is CORS.

The model answers, but the plugin says the answer is unusable

Symptom: “The endpoint answered, but not in a usable shape — the model ignored the requested structure. Try a stronger model.” Or, for an agent: “The tool ran and answered, but not with the structured result the plugin needs.”

Every backend must return a structured result: findings with verbatim quotes, a critique, optional structured edits, a severity. A malformed envelope is still a failure — but a single malformed finding inside a valid review no longer fails the run: its proposal is removed (the card says so) or the finding is dropped, and the panel counts the loss.

A model that wraps its JSON in prose (“Here are my comments: …”) or a code fence is tolerated: the plugin recovers the object and validates it strictly afterwards. What you see as a failure is what could not be recovered.

If the message says the model ran out of output space (or “stopped mid-answer”): the answer hit the model’s output limit before it finished — this is not a formatting problem. A long selection or note produces a long answer. Try a shorter selection, split the note, or use a model with a larger output limit.

• Use a stronger model. Small local models routinely invent fields or drift from the schema. This is the single most common cause.
• Check you are on a chat/instruct model, not a base completion model.
• For agents, expect it more often: an agent that narrates before answering looks exactly like this.
• Simplify the request while diagnosing: one editor, a short selection, thinking off.
• Suspect your extra request body if you set one — a wrong response_format or a routing preference sending you to a different model will do this.

The response is also refused if the envelope exceeds the safety bounds: more than 200 findings, or a replacement over 100000 characters. Per-finding bound violations (an oversized quote or critique, an invalid edit) degrade just that finding instead.

Failures, retries, and what the plugin does on its own

Transient failures are retried automatically, a bounded number of times: a network blip or a 5xx gets a couple of attempts with a short backoff, a timeout and an unusable answer get one more attempt, a rate limit waits the time the provider asked for (up to 30 seconds). What is never retried: a rejected key (fix it in the Backends tab), an exhausted quota or credit (the message says so — paying is the fix, retrying is not), a cancelled run, and an answer that ran out of output space. A backend that keeps failing stops receiving automatic retries until a request succeeds, you run Test connection, or you change settings. Daemon mode additionally turns itself off after three fully-failed automatic refreshes.

Findings appear under “Not anchored”

The editor’s quote could not be located in your text — not exactly, and not uniquely enough to be safe.

That is by design: a finding is only actionable when its quote matches your text exactly, or matches uniquely once whitespace is normalized. Anything vaguer is shown as display-only so you can read the critique, but it is never applied and never guessed into a position.

Usual causes:

• The model paraphrased instead of quoting. Weaker models do this constantly. A stronger model fixes it.
• You edited the text while the review was running. Re-run the review.
• The quote spans a formatting boundary the model reproduced differently — smart quotes, a soft line break, a list marker.

You can still dismiss an unanchored finding; you cannot accept it.

A finding went stale

Dashed and dimmed: you edited inside the highlighted span, so the proposal no longer matches the text it was computed against. Accepting it would apply its edits to something else.

Dismiss it, or run the review again for a fresh look at the text as it now reads.

Review refuses to start

Hover the panel’s Review button — the tooltip says which of these it is.

Reason	Fix
No note open	Open a markdown note
Excluded	The note matches a privacy exclusion — folder, tag, or ai_editor: false
Rule-disabled	A binding rule switched the plugin off for it. The message names the rule
No editor can review	Every editor is disabled, has no usable backend, or has review capability off

A rule-disabled note has no rail, no menu items and no commands either — that is what the kill switch means.

An editor is skipped

Skips are always reported, never silent. The reason is one of:

Message	Fix
review capability disabled	Editors tab → the editor → Review
rewrite capability disabled	Editors tab → the editor → Rewrite (transform actions need it)
no backend configured	Set a global default backend, or assign one to the editor
its backend no longer exists	Reassign the editor to a backend that exists
its backend is disabled	Backends tab → enable it
its CLI backend has not been allowed to run	Grant launch consent on the backend’s row
no model configured	Set a default model on the backend, or a model override on the editor
the editor is disabled	Editors tab → enable it
the matching rule’s panel no longer exists	Rules tab → repoint or delete the rule

The note is too big

Notes above Size warning threshold (default 8000 words) ask for confirmation before anything is sent. Confirm, or:

• review a selection instead of the whole note;
• raise the threshold in the Behavior tab if you routinely work with long notes and know what it costs;
• from the CLI, pass --confirm-large.

Daemon mode silently skips oversized notes rather than interrupting you with a dialog. If a note never refreshes automatically, its size is the first thing to check.

Separately, the context budget (default 200000 characters) governs what fits in one request. The system prompt and the reviewed note are never truncated; attachments are dropped from the end. Preview what will be sent shows exactly what was dropped and why.

A CLI backend will not run

Message	What it means
No executable path is configured	Fill in Executable
The executable path must be absolute	No ~, no bare name, no relative path — no shell runs here to expand them
No file exists at …	Wrong path. which claude / which codex prints the real one
… is not a file	You pointed at a directory
… is not executable by this user	chmod +x, or point at the real binary
This file is a script, not a program	A Windows .cmd/.bat/.ps1 shim. Point at the .exe or use an API backend
has not been allowed to run	Grant the first consent. Enabled is not consent

Detect found nothing? It only probes a curated list of install locations and never searches PATH. Run which claude in a terminal and paste the result.

Consent stopped applying? Consent records which executable it was granted for. Changing the path, importing settings, or syncing data.json from another machine invalidates it, and you are asked again about the program that is actually there. That is deliberate.

The run failed with “the process tree could not be stopped”. An agent left something behind. The run is reported as failed rather than as a success, on purpose. Check for orphaned processes before re-running.

You cannot see the tool’s error text. By design: an agent CLI echoes its configuration when it fails, and configuration contains credentials. You get the status, the byte count of the error stream, and whether something could not be stopped.

Margin comments

The column is not showing. In order of likelihood:

• the pane is narrower than ~700px — the AI Editor Review panel is the surface below that;
• you are in Reading view, which has no column. The comments still exist and are still listed in the panel;
• Margin comment column is off in the Behavior tab, or you hit Toggle the margin comment column.

A comment says it is orphaned. You edited away the text it quoted. The comment is kept with its original quote in a collapsed group at the top of the column rather than deleted. Retry refuses an orphaned comment: re-asking about text that no longer exists would answer a different question.

A comment is stuck as interrupted. Obsidian closed, or the plugin unloaded, while it was in flight. Press Retry — nothing is ever resumed automatically.

A comment never starts. Background work never queues ahead of a review: it takes a slot only when the queue is empty and there is a spare one. On a busy vault a comment can wait. Let the reviews finish, or raise Max concurrent requests.

You were told the store could not be written. New comments are refused for the session rather than accepted and lost at quit. Check that Obsidian can write to its config folder and restart.

Actions

An action is missing from the menu. Unbound actions are hidden. Bound-but-undispatchable ones show the reason under their row in the Actions tab: no name, no instruction, no class chosen, a target that is disabled or deleted.

A custom action does nothing. Check What it does is set. There is deliberately no default, and until you pick one the action stays out of every surface.

A panel-bound action refuses. Only report-findings actions can bind a panel. Changing an action’s class away from report-findings clears the binding.

Panels

A member is missing from the scorecard. It failed, and the panel says so and names it. Retry that member and the scorecard is rewritten.

The scorecard failed but the findings are there. Aggregation is one extra request; when it fails, the member reviews below it are untouched. The block above them says what went wrong.

The scorecard says it is from the previous round. A member is generating more findings, so the scorecard will be rewritten when it finishes. Keeping the old one beats showing nothing.

Commands are missing from the palette

• editor-ai-daemons:* CLI subcommands need Obsidian 1.12.2 or newer — the same floor the plugin itself requires.
• Per-action and per-editor commands are generated from your settings and disappear when the binding does.
• Triage commands only appear when there is something to triage.
• Nothing at all for this note: a binding rule has switched the plugin off for it.

Nothing here helped

Open an issue with what you did, what you expected, and what happened. Include your Obsidian version and which backend kind you use — but never paste your data.json: it contains your API keys.