diff options
Diffstat (limited to 'claude-rules/subagents.md')
| -rw-r--r-- | claude-rules/subagents.md | 57 |
1 files changed, 51 insertions, 6 deletions
diff --git a/claude-rules/subagents.md b/claude-rules/subagents.md index e52d906..b36abba 100644 --- a/claude-rules/subagents.md +++ b/claude-rules/subagents.md @@ -10,7 +10,10 @@ deliberately, not reflexively. ## Pre-Dispatch Checks Run these two checks before any spawn. Both can send the work back to the -main thread without the spawn ever happening. +main thread without the spawn ever happening. The two overrides below +(Isolation, Output Destination) point the other way, at dispatch, and each +lifts the Cost gate. Neither touches Availability: no override can conjure a +spawn mechanism that isn't there. ### Availability @@ -36,7 +39,39 @@ at dispatch time. Every size-based rule in this file — the cost gate here, "Don't Subagent At All", the trivial-work anti-pattern — is subject to the isolation override -below. +and the output-destination gate below. + +## Output-Destination Override: When the Terminal Is the Cost + +The cost gate above weighs the handoff against the work. The isolation +override below fires when the main thread is disqualified from judging. This +third justification is neither. The work is trivial, the main thread is +perfectly capable of it, and dispatching is still correct, because a background +subagent's output goes to a file while the main thread's goes to Craig's +terminal. Background is the operative word. A foreground dispatch puts the +output right back where it started. + +It fires for a **repeated multi-step pass on a schedule**: a recurring battery +of checks, a monitor loop's per-cycle body, any prompt that walks the same list +every N minutes. Ten Bash calls once is fine inline. Ten Bash calls every hour +for eleven hours is a hundred and ten blocks of output in the buffer he's +working in, which is what happened on 2026-08-05. + +Dispatch those, and let the quiet-output rule in `interaction.md` govern what +the main thread then says about the result: deviations, plus one line per cycle. +The justification is where the output lands, not how hard the work is. So the +size gates in this file don't apply, the same way they don't apply under the +isolation override. In particular, "the target is already known and the work +fits in under ~10 tool calls" is exactly the shape of a recurring check battery, +so that gate would otherwise refuse every case this one exists to catch. + +The inverse keeps this from becoming a licence to dispatch everything. Work that +produces little output, or output Craig actually wants to watch, gains nothing +from the handoff and still pays the contract cost. A one-off check stays inline +and prints quietly. + +Craig's directive, 2026-08-05: quiet form always, and hand a repeated pass to a +subagent. ## Isolation Override — When Size Doesn't Gate @@ -123,8 +158,13 @@ to the user to adjudicate, not back to the author's own judgment. ### Don't Subagent At All -Unless the Isolation Override applies — these are efficiency rules, and they -lapse when the main thread's own context is what makes its answer untrustworthy. +Unless one of the two overrides applies. These are efficiency rules, and they +lapse under the Isolation Override (the main thread's own context is what makes +its answer untrustworthy) and under the Output-Destination Override (the pass +repeats on a schedule, so its output floods Craig's terminal). The first bullet +below is the one the second override most often lifts. A recurring check battery +is always a known target in under ~10 calls, and that's not a reason to run it +inline. - **The target is already known** and the work fits in under ~10 tool calls. - **Single-function logic** — one Read + one Edit is faster than briefing @@ -202,8 +242,10 @@ fix), then dispatch the fix with a specific contract. - **Retrying a failed subagent task in the orchestrator** — pollutes context. Dispatch a fix agent instead. - **Subagenting trivial work** — one Read + one Edit doesn't need an - agent; spawn overhead exceeds benefit. Except under the Isolation - Override, where a one-line diff still gets its own reviewer. + agent; spawn overhead exceeds benefit. Except under either override: the + Isolation Override, where a one-line diff still gets its own reviewer, and + the Output-Destination Override, where a trivial pass that repeats hourly + gets dispatched for where its output lands. - **Reviewing your own change inline** — the mirror-image failure, and the more expensive one. Skipping a dispatch to save overhead on a small diff costs a review that could only have come from outside your context. @@ -219,3 +261,6 @@ fix), then dispatch the fix with a specific contract. see `verification.md`. - Testing discipline applies to subagent-produced tests too — see `testing.md`. +- What the main thread prints once a pass is dispatched: see the + quiet-output rule in `interaction.md`. It's the other half of the + Output-Destination Override above. |