Add coding-loop workflow + fix tmux agent stage-status env forwarding

[mattleaverton]

Summary

• New workflow package workflows/coding-loop/ — an iterative coding-agent loop: task chooser → implementer → reviewer → done-gate, wrapped in the trapezium/invtrapezium loop primitive with file-based termination. One sub-task per iteration; feedback persisted to .reviews/iter-NNN.md + rolling .reviews/latest.md; done-gate uses LLM judgment against the spec.
• Engine fix — the tmux agent session now receives KILROY_STAGE_STATUS_PATH and KILROY_STAGE_STATUS_FALLBACK_PATH (plus the full BuildStageRuntimeEnv set). The engine's status-contract preamble already tells agents to write to these vars, but previously only the API agent_loop path actually set them in the process env. The tmux path didn't. This affects every agent_tool=claude|codex|gemini|opencode node in every workflow.

Motivation

Writing and running the coding-loop workflow exposed a latent tmux/API parity gap. When I ran the workflow end-to-end on a toy-list spec (7 sub-tasks → 7 iterations, full green run), I observed the implementer burning roughly 15 of 45 tool calls per iteration hunting for $KILROY_STAGE_STATUS_PATH that was never set:

echo "$KILROY_STAGE_STATUS_PATH"     → (empty)
printenv | grep -i kilroy            → only KILROY_RUN_ID, KILROY_NODE_ID
grep -r "status_path" <logs_root>    → no match in any config

Agents eventually gave up and wrote status.json to arbitrary fallback paths. Runs still succeeded because the engine already tolerates a missing status file, but the wasted latency per iteration was substantial (~2-3 minutes each) and the behavior was confusing (preamble tells you to write to an env var that doesn't exist).

What changed

internal/attractor/agents/tmux_handler.go — extracted the session-env construction into buildTmuxAgentEnv, which now merges:

• The tool template's BuildEnv() defaults (unchanged)
• engine.BuildStageRuntimeEnv — run/node IDs, worktree/logs paths, data dir, inputs manifest, KILROY_INPUT_* (replaces the previous hand-rolled subset)
• engine.BuildStageStatusContract(...).EnvVars — the two status-contract paths (new)

This brings the tmux session env into parity with what buildAgentLoopOverrides already provides to the API agent_loop path.

internal/attractor/agents/tmux_env_test.go — new unit tests covering both populated and nil-template code paths; asserts all expected runtime + status-contract vars are present.

workflows/coding-loop/ — new workflow package (graph.dot, workflow.toml, README.md) exercising the loop primitive end-to-end. Proven on a toy-math spec (1 iter, 4m 24s) and a toy-list spec (7 iters, 30m 54s). Both green.

Test plan

• go test ./internal/attractor/agents/ ./internal/attractor/engine/ — green (agents 7.4s, engine 220.4s)
• kilroy attractor validate --graph workflows/coding-loop/graph.dot — ok
• End-to-end run: toy-math (3 sub-tasks) — 1 iteration, success
• End-to-end run: toy-list (7 sub-tasks, empty-safe invariant) — 7 iterations, all features + tests implemented, go test ./... exits 0 in target repo
• go build ./cmd/kilroy/ — clean
• go vet ./... — clean
• gofmt -l on touched files — clean (pre-existing drift elsewhere unchanged; PR fix(ci): gofmt all unformatted files (engine.go, worktree_hint_test.go, cli_only_models_test.go, codergen_router_cxdb_test.go) #74 covers some of it)

Notes

• Two pre-existing test failures on main (TestRunWithConfig_ForceModel_BypassesCatalogGate, TestRunWithConfig_AllowsKimiAndZai_WhenCatalogUsesOpenRouterPrefixes) are unrelated to this change.
• The workflow package is functional but v0.1.0 — future work could use the housekeeping-LLM primitive (separate exploration) to replace the exact-string termination match, drop the loop_max=12 cap, and make the done-gate more robust to prompt variation.

🤖 Generated with Claude Code