huskies: merge 961

This commit is contained in:
dave
2026-05-13 11:22:57 +00:00
parent 78b1ecdc3c
commit 8b53e20ca9
38 changed files with 327 additions and 146 deletions
+23 -10
View File
@@ -78,10 +78,15 @@ impl AgentPool {
// The coder exited with uncommitted content but no commits.
// Check if this is already a second recovery attempt (the
// first recovery respawn also produced no commits).
let recovery_key = format!("{story_id}:commit_recovery_pending");
if crate::db::read_content(&recovery_key).is_some() {
if crate::db::read_content(crate::db::ContentKey::CommitRecoveryPending(
story_id,
))
.is_some()
{
// Second attempt still produced no commits → block.
crate::db::delete_content(&recovery_key);
crate::db::delete_content(crate::db::ContentKey::CommitRecoveryPending(
story_id,
));
slog!(
"[pipeline] Coder '{agent_name}' (commit-recovery respawn) \
still produced no commits for '{story_id}'. Blocking story."
@@ -99,7 +104,10 @@ impl AgentPool {
} else {
// First occurrence: issue a commit-only recovery respawn.
// This does NOT consume a retry_count slot.
crate::db::write_content(&recovery_key, "1");
crate::db::write_content(
crate::db::ContentKey::CommitRecoveryPending(story_id),
"1",
);
slog!(
"[pipeline] Coder '{agent_name}' exited with uncommitted work \
for '{story_id}'. Issuing commit-only recovery respawn."
@@ -125,7 +133,9 @@ impl AgentPool {
}
} else if completion.gates_passed {
// Clear any stale recovery key when the coder succeeds normally.
crate::db::delete_content(&format!("{story_id}:commit_recovery_pending"));
crate::db::delete_content(crate::db::ContentKey::CommitRecoveryPending(
story_id,
));
// Determine effective QA mode for this story.
let qa_mode = {
let item_type = crate::agents::lifecycle::item_type_from_id(story_id);
@@ -183,7 +193,9 @@ impl AgentPool {
} else {
// Clear any stale recovery key when gates fail normally (agent committed
// but the build is broken — treat as a standard retry, not a recovery).
crate::db::delete_content(&format!("{story_id}:commit_recovery_pending"));
crate::db::delete_content(crate::db::ContentKey::CommitRecoveryPending(
story_id,
));
// Bug 645 / 668: Before retry/block, check if the agent left committed
// work AND the agent had a passing run_tests result captured during its
// session. An agent may crash mid-output (e.g. Claude Code CLI PTY write
@@ -195,8 +207,9 @@ impl AgentPool {
// whenever script/test exits 0 inside a story worktree. Consume the
// evidence here so it does not persist to the next agent session.
let has_test_evidence =
crate::db::read_content(&format!("{story_id}:run_tests_ok")).is_some();
crate::db::delete_content(&format!("{story_id}:run_tests_ok"));
crate::db::read_content(crate::db::ContentKey::RunTestsOk(story_id))
.is_some();
crate::db::delete_content(crate::db::ContentKey::RunTestsOk(story_id));
let work_survived = has_test_evidence
&& worktree_path.as_ref().is_some_and(|wt_path| {
crate::agents::gates::worktree_has_committed_work(wt_path)
@@ -258,7 +271,7 @@ impl AgentPool {
// Persist gate_output so the retry spawn can inject it into
// --append-system-prompt (story 881).
crate::db::write_content(
&format!("{story_id}:gate_output"),
crate::db::ContentKey::GateOutput(story_id),
&completion.gate_output,
);
// Increment retry count and check if blocked.
@@ -384,7 +397,7 @@ impl AgentPool {
// Persist gate_output so the retry spawn can inject it into
// --append-system-prompt (story 881).
crate::db::write_content(
&format!("{story_id}:gate_output"),
crate::db::ContentKey::GateOutput(story_id),
&completion.gate_output,
);
if let Some(reason) = should_block_story(story_id, config.max_retries, "qa") {