feat(sei-agent-driver): set the review's reasoning effort per session

[bdchatham]

Summary

SEIDROID_EFFORT sets the reasoning effort for the review's own agent, for example high. Empty leaves the agent spec's own. Today the review runs at the spec's effort, and the seidroid spec sets none.

Why per session

seidroid is the shared platform bot. Setting effort in its spec would change all of its work. The session API takes reasoning_effort at create, and SetReasoningEffort / ClearReasoningEffort move it later (omnigent-go-sdk v0.3.0). That matches how SEIDROID_MODEL already works.

Change

• Config.Effort reads SEIDROID_EFFORT. Work.Effort is three-valued like Work.Model: nil leaves it alone, a value sets it, empty clears it.
• workFor gives the review's work the configured effort. A scout gets nil, because a scout takes its effort from its own bundle.
• Session create sends ReasoningEffort.
• On adopt, reconcileOverrides moves both the model and the effort through one function, reconcileOverride. It replaces reconcileModel. Log messages now carry an override attribute.
• README: a SEIDROID_EFFORT row.
• A session create that the server refuses with a 400 now exits 2 (ExitConfig), not 6. The server refuses an unknown effort that way, and the SDK reports it as ErrInvalidInput, which the driver did not map.

Tests

Test	What it proves
TestCreateCarriesTheConfiguredEffort	Create sends reasoning_effort: high. Fails without the create field: <nil>, want high.
TestAdoptedSessionIsPointedAtThisRunsEffort	Adopt sends one patch that sets the effort and leaves the model alone. Fails without the reconcile entry: sent 0 session patches.
TestARejectedEffortFailsAsConfiguration	A 400 on create exits 2 and is not retried. Fails without the mapping: ExitCode = 6.
TestCreatedScoutSessionCarriesNoModel	Now also asserts a scout session carries no effort.
Existing model tests	Unchanged and passing on the shared reconcile.

Verification

gofmt -l .                     clean
go vet ./...                   clean
go test -race -count=1 ./...   ok (cmd, driver, omni, review)
golangci-lint run ./...        5 findings, all present on main unchanged
govulncheck ./...              No vulnerabilities found
vale README.md                 the new row is clean

Rollout

Takes effect once uci passes SEIDROID_EFFORT (sei-protocol/uci#118, claude-effort input, default high) and pins a driver that reads it. An adopted session whose harness is already up answers that run at its launch effort. The new value applies from the next launch.

🤖 Generated with Claude Code

[cursor]

PR Summary

Medium Risk
Changes session create/adopt and override reconciliation for every review run; misconfiguration or reconcile bugs could affect which model/effort answers a PR, though invalid effort fails fast at create.

Overview
Adds SEIDROID_EFFORT so review runs can override the agent spec’s reasoning effort (e.g. high) per Omnigent session, mirroring SEIDROID_MODEL. The driver loads it into config, attaches a three-valued Work.Effort for the review agent only (scouts keep nil so they use their own bundle), sends reasoning_effort on session create, and on adopt reconciles model and effort together via a shared reconcileOverrides path (replacing model-only reconcile).

Invalid effort at create (server 400 / SDK ErrInvalidInput) now maps to ExitConfig instead of a generic transport failure. README documents the new variable; tests cover create, adopt, clear, scout isolation, and config exit behavior.

^{Reviewed by Cursor Bugbot for commit 1bfe707. Bugbot is set up for automated code reviews on this repo. Configure here.}

[seidroid]

Adds SEIDROID_EFFORT, a per-session reasoning-effort override for the review's own agent. It is set when a session is created and moved when an existing session is adopted, using the same reconcile path as the model override, which this change generalises. Every call matches omnigent-go-sdk v0.3.0: the create field is omitempty, Set/ClearReasoningEffort exist, and SessionResponse.ReasoningEffort is *string. Scouts correctly get nil, and nothing blocks; codex found nothing, and I agree there are no defects, with two minor suggestions below (no REVIEW.md on the base branch; Go is not installed here, so the tests were not run and this rests on reading the PR tree at cf70afb).

Non-blocking

1 finding on the changed lines, as inline comments.

• No test covers the clear branch for effort: an adopted session that already has a reasoning_effort while SEIDROID_EFFORT is empty should get ClearReasoningEffort. The shared reconcileOverride makes this low-risk, but the model override has its own clear test and effort has none.

_{seidroid review · decision approve · session 203bb184497c42bea42e0c97f9771b9a · turn resp_claude_6d9c5a8e4691fa20ce389370e0708410 · item 1211545875ca5a48bcd547b5b5955050}

Findings: 0 blocking | 2 non-blocking | 1 posted inline

[bdchatham]

@seidroid review

[seidroid]

Since the last review, a 400 on session create (the SDK's ErrInvalidInput, which is how the server refuses an unknown effort) now exits 2 (ExitConfig) instead of 6, and the README says so. TestARejectedEffortFailsAsConfiguration pins both the exit code and that the create is not retried, which addresses my earlier finding at config.go:242, so that thread is resolved. One non-blocking test gap is still open; codex found nothing again, and I agree there are no defects (reviewed against the PR tree at 9cace77; Go is not installed here, so the tests were not run).

Non-blocking

• No test covers clearing effort on adopt: when an existing session already has a reasoning_effort and SEIDROID_EFFORT is empty, the driver should call ClearReasoningEffort. The model override has its own clear test; effort relies only on the shared reconcileOverride.

_{seidroid review · decision approve · session 203bb184497c42bea42e0c97f9771b9a · turn resp_claude_5e249e98e98e065764c6095bd85faeb9 · item d3330648f7de509193ce5d999d9dfa4e}

Findings: 0 blocking | 1 non-blocking | 0 posted inline

[bdchatham]

@seidroid review

[seidroid]

Since the last review, a new test covers clearing effort on an adopted session (TestAdoptedSessionsEffortIsClearedWhenNoneIsConfigured), using a generalised adoptFakeServer, which closes the test gap I raised last time; the earlier config.go:242 thread is already resolved and holds. Nothing blocks; I kept codex's one finding that the clear assertion is loose, as a nit, since with an empty want only the clear path can run and the test can only miss a wrong alias value (reviewed against the PR tree at c0d27f4; Go is not installed here, so the tests were not run).

1 nit, not posted on the code

• sei-agent-driver/internal/omni/effort_test.go:90 — From codex, confirmed: the assertion accepts any non-empty reasoning_effort, so it does not check that the clear alias ("default" in omnigent-go-sdk v0.3.0) was the value sent. The model clear test at model_test.go:137 has the same looseness.

_{seidroid review · decision approve · session 203bb184497c42bea42e0c97f9771b9a · turn resp_claude_a69e7c8bf9ec00017efc1208ff56ce0e · item b64a3901e3ea50efa0c26054ef8ed8ff}

Findings: 0 blocking | 0 non-blocking | 0 posted inline