Squad runs lack native Claude Code tool/subagent parity (Explore/Task, WebSearch); DeepSeek routes via aider

[kokevidaurre]

Problem

Agents launched via squads run do not have the same native tool/subagent
suite that an interactive Claude Code session exposes. Two concrete gaps:

1. No built-in subagents. Native Claude Code offers Explore (read-only
   research/search) and Task/general-purpose subagents. Scheduled squad agents
   can't spawn or use them, so any agent whose job is broad research/search has no
   equivalent capability.

2. provider: deepseek routes through aider — a code-editing assistant with
   no web access. A scheduled web-scanning agent run on DeepSeek produced
   no output: aider has no WebSearch, doesn't do autonomous multi-step
   research, and instead asks the operator to "add files to the chat" so it can
   edit them. Run completed in ~48s having done nothing. So any research/
   scanning agent is a no-op on the DeepSeek/aider path.

Tell-tale: the only agents that work today on provider: deepseek are
file/data workers (validate-files, synthesize-files, tidy-files). Anything that
needs the web silently fails.

Impact

• Off-quota execution (the whole point of DeepSeek routing) is unavailable for
  the largest class of agents — the web-research/scanning ones. They must run on
  a Claude model to function, which defeats the quota-relief goal.
• No Explore/Task subagent parity means squad agents can't decompose
  research the way an interactive session can.

Desired

Tool/subagent parity between native Claude Code and squads run:

1. Expose WebSearch (and WebFetch) to squad agents regardless of provider.
2. Make the Explore / Task subagents available to squad agents.
3. Route DeepSeek through the Anthropic-compatible endpoint (carries the full
   tool suite) instead of the aider backend for non-coding agents. Relates to
   the scoped off-quota work (Provider tiering: route low-tier agent roles to API-billed models (DeepSeek v4 / Qwen3) #820).

Acceptance

• A scheduled web-scanning agent run via squads run can call WebSearch and
  produce its artifact — on a Claude model and on the DeepSeek-compat path.
• Squad agents can invoke Explore/Task subagents.

Evidence

Reproduced 2026-06-19: a web-scanning agent on provider: deepseek (model
deepseek-chat) no-op'd via aider; the same agent reverted to claude-haiku-4-5
to function in the interim.

[kokevidaurre]

Update: validated a second, distinct failure mode on the Claude path

Re-ran the same scheduled web-scanning agent on a Claude model (claude-haiku-4-5)
instead of DeepSeek. Result clarifies the gap is two-dimensional:

• Tools ARE present on the Claude path. The agent correctly planned to call
  WebSearch, query the DB, and poll the feeds — so unlike the DeepSeek/aider path,
  the Claude path exposes the web tools.
• But the scheduled/headless run does not execute autonomously. The agent
  emitted its plan and ended its turn with "Should I proceed?", then the run
  terminated without doing the work — no artifact produced. In a headless
  scheduled context there is no operator to approve, so it no-op's.

The squad's approval policy already marks web.search (and memory.update,
commit.push) as auto-approved, but that policy is not honored in the
headless run — the agent still gates on human approval and then exits.

So tool/subagent parity needs to cover autonomous execution too:

• Scheduled/headless agents must run non-interactively, honoring the squad's
  approvals.auto list (don't pause for an approval that will never come).
• Pre-authorized tool classes (web.search, read-only DB, memory.update) should
  execute without prompting in scheduled mode.

Acceptance, refined: a scheduled web-scanning agent run via squads run (or its
cron) completes its scan and writes its artifact unattended — no "should I
proceed?" dead-end.

(Also noted, separate concern: the founder-context digest pre-flight re-runs on
every --force and takes ~6 min — worth a freshness cap.)