impl-maxagents-authority-fix — WorksGood live mirror

Metadata

Status	failed ‖ paused
Created	2026-07-25T20:56:54.917684992+00:00
Started	2026-07-26T06:09:58.433670256+00:00
Completed	2026-07-27T15:48:41.489684274+00:00
Tags	`coordinator`, `config`, `bugfix`
Tokens	5591331 in / 27129 out
Failure reason	required evaluation gate rejected: evaluator verdict-evalp-fd5da61975fbab269deb9f97-evaluate-f3154d11cfa45430 score=0.64 threshold=0.70 FAIL; FLIP verdict-evalp-fd5da61975fbab269deb9f97-flip-935878596618c786 score=0.72 threshold=0.70 PASS

Description

Objective

Fix the max_agents four-source-of-truth authority bug so a runtime override survives a flagless wg service reload / profile swap. This is the SHARED PREREQUISITE for the adaptive-parallelism controller (T4) — without it, the controller's writes are silently clobbered. It is a pure correctness fix valuable independently (fixes the observed "start with --max-agents 2, wg profile use, observe 8" bug).

Context: docs/studies/adaptive-parallelism-budget-design.md §8 (the bug + the fix), docs/studies/roadmap-rate-limit-and-distribution.md §2.2/§3.

What to build

There are FOUR sources of truth for max_agents today with no defined precedence: (1) CLI launch arg --max-agents (daemon memory only, src/commands/service/mod.rs:~2415), (2) [coordinator].max_agents in merged config.toml (src/config.rs:~4089), (3) the active profile overlay (src/profile/named.rs:~473,479, written into config.toml on activation), (4) runtime IPC Reconfigure { max_agents } (src/commands/service/ipc.rs:~1095). The bug: the launch arg (1) is never persisted, so a flagless reload sends Reconfigure { max_agents: None } (src/commands/service/mod.rs:~3891, src/commands/profile_cmd.rs:~1205), handle_reconfigure takes its else-branch, re-reads config.toml, and does daemon_cfg.max_agents = config.coordinator.max_agents (src/commands/service/ipc.rs:~1111) — silently replacing the launch arg's value with the profile's.

Implement the fix exactly as specified in the study §8.2:

Add runtime_max_agents: Option<usize> to CoordinatorState (src/commands/service/mod.rs:~710). This is the new single transient-but-persisted authority. It survives daemon restart (on disk in the coordinator state file).
In handle_reconfigure's else-branch (src/commands/service/ipc.rs:~1109-1113): after re-reading config, if a runtime_max_agents is present in CoordinatorState, KEEP it instead of clobbering with config.coordinator.max_agents.
At daemon start (src/commands/service/mod.rs:~2415): if cli_max_agents is Some(n) AND no runtime_max_agents is already on disk, write runtime_max_agents = n into CoordinatorState (so the launch intent survives reload) and arm a session pin.
config.coordinator.max_agents becomes the ceiling + cold-start default only.
Add a --no-pin escape on service start for tests wanting the old behavior.
An explicit --max-agents flag on reload is a human action and wins (it takes the has_overrides branch and should be recorded as a pin).

No schema migration needed (the field is Option, defaults to None = today behavior).

File scope

src/commands/service/ipc.rs (handle_reconfigure else-branch)
src/commands/service/mod.rs (CoordinatorState field + daemon-start pin logic)
src/config.rs (if a --no-pin flag / config key is needed)
New unit test(s)

Implement directly. Do not decompose further unless you hit genuine context pressure.

Validation

New unit test reproduces the reload-clobber: start with --max-agents 2, simulate a flagless Reconfigure { max_agents: None }, assert daemon_cfg.max_agents == 2 (preserved), NOT 8.
Unit test: an explicit --max-agents reload flag overrides the runtime value and is recorded as a pin.
Unit test: daemon restart restores runtime_max_agents from disk.
--no-pin escape restores today behavior for tests.
cargo build + cargo test pass with no regressions.
cargo fmt --check + cargo clippy clean (CI fast-fails on fmt).
No model pin in the task; routes through the active zai profile.

## Objective
Fix the `max_agents` four-source-of-truth authority bug so a runtime override survives a flagless `wg service reload` / profile swap. This is the SHARED PREREQUISITE for the adaptive-parallelism controller (T4) — without it, the controller's writes are silently clobbered. It is a pure correctness fix valuable independently (fixes the observed "start with --max-agents 2, `wg profile use`, observe 8" bug).

Context: `docs/studies/adaptive-parallelism-budget-design.md` §8 (the bug + the fix), `docs/studies/roadmap-rate-limit-and-distribution.md` §2.2/§3.

## What to build
There are FOUR sources of truth for `max_agents` today with no defined precedence: (1) CLI launch arg `--max-agents` (daemon memory only, `src/commands/service/mod.rs:~2415`), (2) `[coordinator].max_agents` in merged config.toml (`src/config.rs:~4089`), (3) the active profile overlay (`src/profile/named.rs:~473,479`, written into config.toml on activation), (4) runtime IPC `Reconfigure { max_agents }` (`src/commands/service/ipc.rs:~1095`). The bug: the launch arg (1) is never persisted, so a flagless reload sends `Reconfigure { max_agents: None }` (`src/commands/service/mod.rs:~3891`, `src/commands/profile_cmd.rs:~1205`), `handle_reconfigure` takes its else-branch, re-reads config.toml, and does `daemon_cfg.max_agents = config.coordinator.max_agents` (`src/commands/service/ipc.rs:~1111`) — silently replacing the launch arg's value with the profile's.

Implement the fix exactly as specified in the study §8.2:
1. Add `runtime_max_agents: Option<usize>` to `CoordinatorState` (`src/commands/service/mod.rs:~710`). This is the new single transient-but-persisted authority. It survives daemon restart (on disk in the coordinator state file).
2. In `handle_reconfigure`'s else-branch (`src/commands/service/ipc.rs:~1109-1113`): after re-reading config, if a `runtime_max_agents` is present in `CoordinatorState`, KEEP it instead of clobbering with `config.coordinator.max_agents`.
3. At daemon start (`src/commands/service/mod.rs:~2415`): if `cli_max_agents` is `Some(n)` AND no `runtime_max_agents` is already on disk, write `runtime_max_agents = n` into `CoordinatorState` (so the launch intent survives reload) and arm a session pin.
4. `config.coordinator.max_agents` becomes the ceiling + cold-start default only.
5. Add a `--no-pin` escape on `service start` for tests wanting the old behavior.
6. An explicit `--max-agents` flag on `reload` is a human action and wins (it takes the `has_overrides` branch and should be recorded as a pin).

No schema migration needed (the field is `Option`, defaults to `None` = today behavior).

## File scope
- `src/commands/service/ipc.rs` (handle_reconfigure else-branch)
- `src/commands/service/mod.rs` (CoordinatorState field + daemon-start pin logic)
- `src/config.rs` (if a `--no-pin` flag / config key is needed)
- New unit test(s)

Implement directly. Do not decompose further unless you hit genuine context pressure.

## Validation
- [ ] New unit test reproduces the reload-clobber: start with `--max-agents 2`, simulate a flagless `Reconfigure { max_agents: None }`, assert `daemon_cfg.max_agents == 2` (preserved), NOT 8.
- [ ] Unit test: an explicit `--max-agents` reload flag overrides the runtime value and is recorded as a pin.
- [ ] Unit test: daemon restart restores `runtime_max_agents` from disk.
- [ ] `--no-pin` escape restores today behavior for tests.
- [ ] `cargo build` + `cargo test` pass with no regressions.
- [ ] `cargo fmt --check` + `cargo clippy` clean (CI fast-fails on fmt).
- [ ] No model pin in the task; routes through the active zai profile.

Depends on

done synthesis-roadmap-from

Required by

done .flip-impl-maxagents-authority-fix

Log

2026-07-25T20:56:54.894911933+00:00 Task paused
2026-07-25T20:57:27.019239946+00:00 Task published
2026-07-25T21:00:41.329367606+00:00 Spawned by coordinator --executor pi --model zai:glm-5.2 --isolation required-worktree
2026-07-25T21:03:50.415154737+00:00 Plan: (1) add runtime_max_agents: Option<usize> to CoordinatorState, (2) add resolve_startup_max_agents helper for daemon-start precedence, (3) fix handle_reconfigure else-branch to preserve runtime override + record explicit --max-agents pin, (4) add --no-pin flag to service start/daemon, (5) write unit tests for all 4 validation criteria.
2026-07-26T05:53:55.243770533+00:00 Dead attempt contained ENOSPC evidence — skipped quality triage/evaluation, released its cache lease, and queued a safe in-place retry
2026-07-26T05:53:58.460418346+00:00 Minted evaluation pipeline evalp-49bc7518f48cbb642e29a3e9 for source attempt 2 (coordinator dead-agent retry)
2026-07-26T06:09:58.505104247+00:00 Spawned by coordinator --executor pi --model zai:glm-5.2 --isolation required-worktree
2026-07-26T06:42:45.151090503+00:00 Implementation complete: added CoordinatorState.runtime_max_agents (single persisted authority), resolve_startup_max_agents precedence ladder, handle_reconfigure keeps runtime override on flagless reload + records explicit --max-agents as pin, --no-pin escape. Found+fixed a tick-clobber race (long-lived in-memory coord_state was overwriting the disk pin on tick save — now reloads runtime_max_agents from disk like accumulated_tokens). 6 unit tests + 1 smoke scenario (which caught the tick race). Also fixed pre-existing last_spawn_failure_at compile break in integration_auto_assignment.rs (field exists on main, test was missing it).
2026-07-26T06:43:17.292589247+00:00 Committed: 135b70d5 — pushed to origin/wg/agent-808/impl-maxagents-authority-fix
2026-07-26T06:45:40.560051898+00:00 Task pending required evaluation gate (agent reported done; awaiting exact attempt-bound `.flip-*`/`.evaluate-*` verdicts)
2026-07-26T06:45:52.084888227+00:00 Squash-merged to main (eb435fa1). Resolved concurrent merge conflict on grow-only tests/smoke/manifest.toml (kept both my service_max_agents_survives_reload scenario AND T6's release_workflow_signing_contract scenario). VALIDATION COMPLETE: 6 unit tests (4 resolve_startup_max_agents precedence + 2 handle_reconfigure preserve/pin) + 1 smoke scenario all pass; cargo build/test/fmt --check/clippy clean; 353 service + 22 integration + 3 integration_service tests pass, no regressions. Bonus: end-to-end probe found+fixed a tick-clobber race (long-lived in-memory coord_state overwrote the disk pin on tick save) that the unit tests structurally cannot reach — the smoke scenario pins it.
2026-07-26T06:51:10.373098757+00:00 Pinned historical PendingEval as a required gate at evaluator threshold 0.70
2026-07-26T07:15:23.259687967+00:00 Pinned historical PendingEval as a required gate at evaluator threshold 0.70
2026-07-26T07:15:47.003302771+00:00 Pinned historical PendingEval as a required gate at evaluator threshold 0.70
2026-07-26T07:15:54.193556125+00:00 Pinned historical PendingEval as a required gate at evaluator threshold 0.70
2026-07-26T07:16:17.537116360+00:00 Pinned historical PendingEval as a required gate at evaluator threshold 0.70
2026-07-26T07:16:43.433810574+00:00 Pinned historical PendingEval as a required gate at evaluator threshold 0.70
2026-07-26T07:17:06.951188981+00:00 Pinned historical PendingEval as a required gate at evaluator threshold 0.70
2026-07-26T07:17:30.630192193+00:00 Pinned historical PendingEval as a required gate at evaluator threshold 0.70
2026-07-26T07:17:52.796408265+00:00 Pinned historical PendingEval as a required gate at evaluator threshold 0.70
2026-07-26T07:17:57.798067784+00:00 Pinned historical PendingEval as a required gate at evaluator threshold 0.70
2026-07-26T07:18:14.610432481+00:00 Pinned historical PendingEval as a required gate at evaluator threshold 0.70
2026-07-26T07:18:19.991353960+00:00 Pinned historical PendingEval as a required gate at evaluator threshold 0.70
2026-07-26T07:18:39.765667190+00:00 Pinned historical PendingEval as a required gate at evaluator threshold 0.70
2026-07-26T07:18:44.712129157+00:00 Pinned historical PendingEval as a required gate at evaluator threshold 0.70
2026-07-26T07:19:03.247370895+00:00 Pinned historical PendingEval as a required gate at evaluator threshold 0.70
2026-07-26T07:19:08.233926794+00:00 Pinned historical PendingEval as a required gate at evaluator threshold 0.70
2026-07-26T07:19:23.899663699+00:00 Pinned historical PendingEval as a required gate at evaluator threshold 0.70
2026-07-26T07:28:04.759793779+00:00 Pinned historical PendingEval as a required gate at evaluator threshold 0.70
2026-07-26T07:28:09.766819580+00:00 Repaired evaluation pipeline drift for authoritative source attempt 2 (evalp-49bc7518f48cbb642e29a3e9)
2026-07-26T07:57:06.580092997+00:00 Pinned historical PendingEval as a required gate at evaluator threshold 0.70
2026-07-26T08:14:35.961683300+00:00 Task reset for retry from pending-eval/failed-pending-eval (stuck evaluation gate) (attempt #1) — reason: Operator recovery: clear legacy pending-eval hold with the newly sanctioned CLI; keep paused until reliability review.
2026-07-26T08:14:35.961692060+00:00 Cleared stuck evaluation gate via `wg retry` — a fresh attempt will be minted; no graph.jsonl edit required.
2026-07-26T08:14:35.961720411+00:00 Minted evaluation pipeline evalp-fd5da61975fbab269deb9f97 for source attempt 3 (resume-in-place retry)
2026-07-26T08:14:36.014993597+00:00 Task paused
2026-07-27T15:45:16.022024481+00:00 Task pending required evaluation gate (agent reported done; awaiting exact attempt-bound `.flip-*`/`.evaluate-*` verdicts)
2026-07-27T15:48:41.489684454+00:00 Consumed durable verdict verdict-evalp-fd5da61975fbab269deb9f97-evaluate-f3154d11cfa45430 exactly once under strict required-gate policy: evaluator verdict-evalp-fd5da61975fbab269deb9f97-evaluate-f3154d11cfa45430 score=0.64 threshold=0.70 FAIL; FLIP verdict-evalp-fd5da61975fbab269deb9f97-flip-935878596618c786 score=0.72 threshold=0.70 PASS; outcome=failed