Delodocs Watchdog Recovery

DeLoDocs Watchdog Recovery

Use this when the vault automation looks superficially healthy but the Inbox is not draining:

• delodocs-inbox-watchdog.service is active
• router logs keep saying pipeline-status is new
• watchdog state shows many files at status: failed with attempts: 3
• a provider/model/key fix was already applied, but stuck files still do not retry

This skill is specifically for the recovery path after root auth/model problems are fixed.

Trigger signals

Reach for this skill when any of these are true:

1. Current Inbox/*.md files still contain pipeline-status: [new]
2. ~/.local/state/delodocs/inbox-watchdog/state.json shows current Inbox files with status: failed
3. Resetting state on disk seems to work temporarily, then reverts back to stale failures
4. A one-shot watchdog run says another watchdog is already running because the service lock file exists

Core diagnosis

The watchdog is stateful in two places at once:

• on disk: ~/.local/state/delodocs/inbox-watchdog/state.json
• in memory: the running watchdog process

The dangerous failure mode is:

1. old watchdog process loads stale failed state into memory
2. operator edits state.json on disk to reset stuck files
3. running watchdog continues with its old in-memory copy
4. running watchdog writes the stale failure state back to disk
5. it looks like the reset "didn't stick"

This is why a reset must be coordinated with service lifecycle.

Recovery rule

Do NOT reset stuck entries while the watchdog service is still running. Stop it first, repair state, run a backfill, then start it again.

Known-good recovery sequence

1. Stop the service

systemctl --user stop delodocs-inbox-watchdog.service

2. Back up state first

cp ~/.local/state/delodocs/inbox-watchdog/state.json \
  ~/.local/state/delodocs/inbox-watchdog/state.json.bak-$(date -u +%Y%m%dT%H%M%SZ)

3. Reset only the current stuck Inbox files

Target: current Inbox/*.md files that still have pipeline-status: [new].

For each such file entry in state.json:

• set status to pending
• set attempts to 0
• update size, mtime_ns, sha256, last_seen_at, last_changed_at
• clear these fields if present:
  • processed_sha256
  • processed_at
  • last_error
  • last_attempt_at
  • enrichment_policy_version
  • enrichment_policy_sha256
  • enrichment_policy_at
  • last_output
  • missing_at

If a current Inbox file is absent from watchdog state entirely, create a fresh entry with:

• first_seen_at
• last_seen_at
• last_changed_at
• size
• mtime_ns
• sha256
• status: pending
• attempts: 0

Why all this clearing matters:

• clearing only attempts is not enough
• if processed_sha256 == sha256, the watchdog can still skip the file

4. Run a one-shot backfill while the service is stopped

python3 _vault/scripts/inbox_watchdog.py \
  --vault "$PWD" \
  --state "$HOME/.local/state/delodocs/inbox-watchdog/state.json" \
  --scan-existing \
  --once

Useful tweak during manual recovery:

python3 _vault/scripts/inbox_watchdog.py \
  --vault "$PWD" \
  --state "$HOME/.local/state/delodocs/inbox-watchdog/state.json" \
  --scan-existing \
  --debounce 0 \
  --once

5. Start the service again

systemctl --user start delodocs-inbox-watchdog.service

6. Verify progress

Check that:

• current Inbox files are no longer all failed
• router logs stop accumulating pipeline-status is new holds for those files
• watchdog logs show actual processing rather than only no ready files

OpenRouter auth source for this vault

The intended runtime for the watchdog is:

• provider: openrouter
• model: deepseek/deepseek-v4-flash
• key source: opg openrouter hermes

Important pitfall:

• opg is a shell function, not a standalone binary
• non-interactive contexts must source the helper file first

Known-good pattern:

zsh -fc 'source ~/.config/zshyzsh/helpers.zsh; opg openrouter hermes'

This matters for repo-local env files or shell startup files that export OPENROUTER_API_KEY.

Verification checklist

After recovery, confirm all of the following before declaring success:

1. systemctl --user show delodocs-inbox-watchdog.service -p ExecMainPID -p ActiveEnterTimestamp --value shows a fresh process if you restarted it
2. current Inbox file status counts in watchdog state are no longer dominated by failed
3. router state/logs show movement or at least fewer held entries caused by pipeline-status is new
4. a direct Hermes OpenRouter call works with the repo-local config

Relationship to other skills

• Use delodocs-vault-pipeline for overall pipeline architecture and first-pass diagnosis
• Use this skill for the specific recovery path where stale watchdog state survives a provider/auth fix