External/buzz

Fork 0

mirror of https://github.com/block/buzz.git synced 2026-08-18 06:50:31 +02:00

Files

T

History

0c33a8a55f fix(agents): canonicalize stale persona harness pins (#4631 )

Replace the stale `agent_command_override` drop logic in
`apply_persona_snapshot` with a three-tier canonical command resolver.

## What this fixes

The old code dropped a create-time harness pin when the persona switched
to a different runtime, but it had two failure modes:

1. **Preset harnesses invisible.** `known_acp_runtime_exact()` only
searches `KNOWN_ACP_RUNTIMES` (builtins). Preset harnesses such as
OpenClaw live in `PRESET_HARNESSES`, so the destination lookup returned
`None` and the outer `if let` branch never executed — a Goose→OpenClaw
persona switch left the stale Goose override in place, keeping the agent
running Goose instead of OpenClaw.

2. **Pin-side canonical resolution incomplete.** The pin was resolved by
`known_acp_runtime()`, which searches by id/command/alias and returns a
`&KnownAcpRuntime` entry correctly. However, if the *pin* named an alias
(e.g. `claude-code-acp`) and the *destination* was a preset harness
absent from builtins, the outer guard still failed for the same reason
as (1). The alias regression test pins the requirement that the
canonical resolver must handle both sides: alias pins must be recognised
and drops must fire when the destination is a known preset.

## How it works now

`canonical_harness_command(input)` accepts any form a stored override
can take — bare command, alias, path prefix, or runtime id — and
resolves it to the harness primary command through three tiers:

1. **Builtins** — `KNOWN_ACP_RUNTIMES`, matched by id/command/alias.
2. **Static presets** — `PRESET_HARNESSES`, matched by id or normalised
command.
3. **Loaded registry** — custom/preset definitions loaded at runtime.

`command_for_runtime_id` (id-only input, same three tiers) replaces the
two-step `known_acp_runtime_exact`/`lookup_loaded_harness_by_id` pattern
in `record_agent_command`, `effective_agent_command`, and
`try_record_agent_command`, adding the static preset tier so preset
harnesses resolve correctly even without a warm registry.

## Changed files

- `discovery/presets.rs` — `preset_command_for_id`,
`command_for_runtime_id`, `canonical_harness_command`
- `discovery.rs` — re-export new functions; make
`normalize_command_identity` `pub(crate)`; refactor three
command-resolution functions to use `command_for_runtime_id`
- `custom_harnesses.rs` — `loaded_harness_registry` visibility `fn` →
`pub(super)` (needed by `canonical_harness_command`)
- `persona_events.rs` — replace two-step
`known_acp_runtime_exact`/`known_acp_runtime` + pointer comparison with
canonical-command comparison
- `persona_events/stale_pin_tests.rs` (new) — four regression tests:
Goose→OpenClaw drop, OpenClaw→Goose drop, claude-code-acp alias→OpenClaw
drop, same-harness path keep
- `persona_events/tests.rs` — `sample_record`/`sample_persona` exposed
as `pub(super)` for the new test module

Signed-off-by: Will Pfleger <pfleger.will@gmail.com>
Co-authored-by: npub1g8493u0xfsjrvflg4n08ezd7vec99mnwzlv0qgwpr9d7gvjwhuzqx59rhw <41ea58f1e64c243627e8acde7c89be667052ee6e17d8f021c1195be4324ebf04@buzz.block.builderlab.xyz>

2026-08-04 13:52:35 -04:00