agency-picks-claude — Workgraph live mirror

Metadata

Status	done
Assigned	`agent-164`
Agent identity	`3184716484e6f0ea08bb13539daf07686ee79d440505f1fdf2de0357707034c3`
Created	2026-04-26T23:08:19.393268063+00:00
Started	2026-04-26T23:13:03.152880045+00:00
Completed	2026-04-26T23:59:05.300956648+00:00
Tags	`eval-scheduled`
Eval score	0.91
└ blocking impact	0.95
└ completeness	0.95
└ constraint fidelity	0.85
└ coordination overhead	0.95
└ correctness	0.95
└ downstream usability	0.90
└ efficiency	0.85
└ intent fidelity	0.88
└ style adherence	0.90

Description

In ~/autohaiku, dispatcher's SpawnPlan log shows:

[dispatcher] implement-haiku-generation-module: SpawnPlan executor=claude (from agency.effective_executor), model=local:qwen3-coder (from dispatcher.default_model), endpoint=none (none (executor=claude))

Then claude CLI gets --model qwen3-coder (after stripping the local: prefix) and returns 404: 'There's an issue with the selected model (qwen3-coder). It may not exist or you may not have access to it.'

Agency chose executor=claude but the model is local:qwen3-coder — that combo CANNOT WORK. claude CLI doesn't know qwen3-coder; it only knows opus/sonnet/haiku and Anthropic-hosted variants. The model's local: prefix says explicitly 'this model lives at a local OAI-compat endpoint, not Anthropic.' Routing it through claude CLI = guaranteed 404.

100% failure rate on autohaiku tasks because of this. SpawnPlan is doing its job (honoring agency.effective_executor as the floor) — the bug is in the agency's executor selection: it should never pick claude executor for a model with the local: prefix.

Fix

In the agency's effective_executor resolution (likely src/agency/ or wherever 'agency.effective_executor' is computed for SpawnPlan):

If the model spec has prefix local: or openrouter: or oai-compat: (anything non-claude/non-anthropic), the agency CANNOT pick claude executor regardless of role/tradeoff defaults. Force a compatible executor (native for local/openrouter, etc).
If the agency's role/tradeoff implies claude executor BUT the model is incompatible, log a clear conflict warning and either (a) override executor to compatible, OR (b) override model to a claude-compatible default (per dispatcher tier config).

Option (a) is safer — preserves the user's model choice.

Add a clear log line on the conflict resolution: 'Agency picked executor=claude for role X but model local:qwen3-coder requires native; routed to native.'

Hard gate

Before claiming done:

In a scratch dir: wg init -m qwen3-coder -e https://lambda01.tail334fe6.ts.net:30000 -x nex; wg service start
wg add 'simple test task' -d 'echo hello world'
wg publish simple-test-task
Wait for dispatch
Daemon log MUST show SpawnPlan with executor=native (not claude); agent metadata MUST show executor=native; task MUST succeed (or fail for a non-routing reason)

Validation

Failing test first: test_agency_does_not_pick_claude_for_local_model — given a model spec with local: prefix and an agency role that defaults to claude, effective_executor returns native (not claude)
Implementation: agency executor selection respects model-prefix compatibility constraints
cargo build + cargo test pass with no regressions
HARD GATE manual smoke as above

## Description

In ~/autohaiku, dispatcher's SpawnPlan log shows:
```
[dispatcher] implement-haiku-generation-module: SpawnPlan executor=claude (from agency.effective_executor), model=local:qwen3-coder (from dispatcher.default_model), endpoint=none (none (executor=claude))
```

Then claude CLI gets `--model qwen3-coder` (after stripping the `local:` prefix) and returns 404: 'There's an issue with the selected model (qwen3-coder). It may not exist or you may not have access to it.'

Agency chose executor=claude but the model is local:qwen3-coder — that combo CANNOT WORK. claude CLI doesn't know qwen3-coder; it only knows opus/sonnet/haiku and Anthropic-hosted variants. The model's `local:` prefix says explicitly 'this model lives at a local OAI-compat endpoint, not Anthropic.' Routing it through claude CLI = guaranteed 404.

100% failure rate on autohaiku tasks because of this. SpawnPlan is doing its job (honoring agency.effective_executor as the floor) — the bug is in the agency's executor selection: it should never pick claude executor for a model with the local: prefix.

### Fix

In the agency's effective_executor resolution (likely src/agency/ or wherever 'agency.effective_executor' is computed for SpawnPlan):

- If the model spec has prefix `local:` or `openrouter:` or `oai-compat:` (anything non-claude/non-anthropic), the agency CANNOT pick claude executor regardless of role/tradeoff defaults. Force a compatible executor (native for local/openrouter, etc).
- If the agency's role/tradeoff implies claude executor BUT the model is incompatible, log a clear conflict warning and either (a) override executor to compatible, OR (b) override model to a claude-compatible default (per dispatcher tier config).

Option (a) is safer — preserves the user's model choice.

Add a clear log line on the conflict resolution: 'Agency picked executor=claude for role X but model local:qwen3-coder requires native; routed to native.'

### Hard gate

Before claiming done:
1. In a scratch dir: `wg init -m qwen3-coder -e https://lambda01.tail334fe6.ts.net:30000 -x nex; wg service start`
2. `wg add 'simple test task' -d 'echo hello world'`
3. `wg publish simple-test-task`
4. Wait for dispatch
5. Daemon log MUST show SpawnPlan with executor=native (not claude); agent metadata MUST show executor=native; task MUST succeed (or fail for a non-routing reason)

## Validation

- [ ] Failing test first: test_agency_does_not_pick_claude_for_local_model — given a model spec with `local:` prefix and an agency role that defaults to claude, effective_executor returns native (not claude)
- [ ] Implementation: agency executor selection respects model-prefix compatibility constraints
- [ ] cargo build + cargo test pass with no regressions
- [ ] HARD GATE manual smoke as above

Depends on

done .assign-agency-picks-claude

Required by

(none)

Log

2026-04-26T23:08:19.390098798+00:00 Task paused
2026-04-26T23:08:19.880993787+00:00 Task published
2026-04-26T23:13:03.152884824+00:00 Spawned by coordinator --executor claude --model opus
2026-04-26T23:13:20.385913630+00:00 Starting investigation: looking for agency.effective_executor resolution
2026-04-26T23:30:02.869541752+00:00 Implemented Agent::effective_executor_for_model(model) — overrides claude → native when model has non-Anthropic provider prefix. Wired into dispatcher (coordinator.rs) and IPC spawn (ipc.rs).
2026-04-26T23:30:06.038869451+00:00 Validated: 3 new unit tests pass (test_agency_does_not_pick_claude_for_local_model, test_agency_keeps_claude_for_anthropic_model, test_agency_does_not_override_explicit_non_claude_executor); all 198 agency tests pass; full --lib suite 1950/1950 pass; --tests suite 2840/2841 pass with 1 pre-existing failure (provenance_full_lifecycle_all_ops_recorded — reproduced on clean tree, noted in wave-1-integration-smoke)
2026-04-26T23:50:04.054994935+00:00 Hard gate PASS: smoke scenario agency_local_model_overrides_claude_executor — daemon SpawnPlan log shows 'executor=native (from agency.effective_executor)' for a default-claude agent + model=local:qwen3-coder, exactly the autohaiku case. Pre-fix output was 'executor=claude'.
2026-04-26T23:50:38.744071426+00:00 Committed: b85ed060e — pushed to remote
2026-04-26T23:59:05.300962940+00:00 Task marked as done