fix(continuation): #987-completion — reset chain-budget for ordinary subagent-returns (#989-P2-1)#992
Conversation
…989-P2-1) #989 reset the chain budget at turn-entry gated on `!isContinuationWake`, so a genuine external turn resets while a continuation-wake preserves the count. But subagent-announce minted `continuationTrigger: "delegate-return"` for EVERY inter-session subagent completion, not just actual `[continuation:chain-hop:N]` continuation-chain hops. So the reset gate treated every ordinary subagent return as an in-chain wake and skipped the reset — a long-lived session with a stale at-cap `continuationChainCount` then rejected any continuation elected from a normal subagent return (the #987 n/200 doom-lock, still open via the ordinary-subagent-return path). Distinguish the two at the source: an in-chain chain-hop return keeps `delegate-return` (mid-chain wake, preserves the leash); an ordinary subagent completion now carries the new `subagent-return` trigger, which get-reply-run classifies as a non-wake (`isContinuationWake=false`) so the reset gate rewinds the chain budget like any other external turn-entry. `subagent-return` stays heartbeat-equivalent and `continuation-chain` fire-reason in run-provenance, so only the reset-gate behavior changes. The runaway leash for genuine continue_delegate chains is preserved. Tests: get-reply-run maps subagent-return -> not a continuation wake; subagent-announce mints subagent-return for ordinary completions and delegate-return for chain hops; run-provenance preserves provenance; the #987 reset-gate tests stay green plus a new #989 at-cap doom-lock reset scenario. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 43ed057d29
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| ? isContinuationChainDelegate | ||
| ? "delegate-return" | ||
| : "subagent-return" |
There was a problem hiding this comment.
Require trusted chain metadata before preserving budget
When an ordinary subagent task text happens to contain the internal marker [continuation:chain-hop:N] (for example, a user asks a subagent to inspect docs/logs containing that literal), this branch now tags its completion as delegate-return, so get-reply-run treats it as a continuation wake and skips the chain-budget reset. I checked the actual chain-hop spawn sites (agent-runner.ts, delegate-dispatch.ts, and the recursive spawn paths in this module) and they also set drainsContinuationDelegateQueue: true, so relying only on the untrusted task string leaves a residual #989 doom-lock path for ordinary completions; require trusted chain metadata as well before emitting delegate-return.
Useful? React with 👍 / 👎.
| const delegateReturnTrigger: ContinuationTrigger | undefined = continuationEnabled | ||
| ? isContinuationChainDelegate | ||
| ? "delegate-return" | ||
| : "subagent-return" | ||
| : undefined; |
There was a problem hiding this comment.
Split silent return wakes before the direct path
This new subagent-return trigger is only passed to deliverSubagentAnnouncement, but the silent/targeted return branches above this block return before reaching it; for example an ordinary silentAnnounce + wakeOnReturn subagent still calls requestHeartbeatNow(... reason: "silent-wake-enrichment"), which heartbeat-runner.ts maps back to continuationTrigger: "delegate-return". Because silentAnnounce is a general spawn option for ambient enrichment and is not restricted to [continuation:chain-hop:N] tasks, those ordinary silent-wake completions still skip the reset gate and can reproduce the stale chain-budget doom-lock this change is meant to close.
Useful? React with 👍 / 👎.
| ? isContinuationChainDelegate | ||
| ? "delegate-return" | ||
| : "subagent-return" |
There was a problem hiding this comment.
Update ordinary completion tests for the new trigger
This value changed the ordinary direct-announce path from delegate-return to subagent-return, but existing ordinary-completion coverage in src/agents/subagent-announce.format.e2e.test.ts still asserts continuationTrigger: "delegate-return" for non-chain-hop completions. In a checkout with dependencies this focused suite will fail even when the runtime behavior is correct, so the commit needs to update those expectations alongside the trigger split.
Useful? React with 👍 / 👎.
a437ca7
into
frond-scribe/20260609/assembly-token-wiring
Completes the #987 doom-lock cure (#989, merged). Closes the residual Codex P2 on #989: figs's n/200 "195-forever" can still bite via the ordinary-subagent-return path.
The hole
#989's reset gate is gated on
!isContinuationWake, preserving the chain-budget on continuation-wakes. Butsubagent-announce.tssetcontinuationTrigger: "delegate-return"for every subagent completion — including ordinary inter-session subagent calls that are NOT in-chain continuation hops. So the gate treated an ordinary subagent-return as an in-chain wake → skipped the reset → a long-lived session's stalecontinuationChainCountpersisted → a continuation elected from that fresh completion was rejected against the old cap. The #987 fix was incomplete for the non-continuation external turn that arrives as an ordinary subagent-return.The fix (distinguish the two return-kinds at the trigger)
Mint a new
subagent-returntrigger for ordinary inter-session subagent completions, distinct fromdelegate-return(kept for in-chain[continuation:chain-hop:N]returns):get-reply-run→ the continuation: chain-count + token-cost caps never reset on normal turns (perma-accumulate bug; reset on fresh non-continuation turn-entry) #987 reset gate rewinds the chain-budget (cures the n/200 hole via this path).delegate-return→ unchanged → preserves + caps (the runaway-leash stays intact for genuine self-continuation chains).run-provenancekeepssubagent-returnheartbeat-equivalent + continuation-chain fire-reason — only the reset-gate classification changes. Protocol enum (gateway-protocol) + RFC doc (continue-work-signal-v2.md) updated.Diff
11 files, +286/-83:
subagent-announce.ts(+test),get-reply-run.ts(+media-only test),get-reply-options.types.ts,agent-runner.ts(+work-span test),run-provenance.ts(+test),gateway-protocol/schema/agent.ts(enum),continue-work-signal-v2.md(RFC).Gate status
test-projects.mjs): 83 pass / 5 fail — all in the known pre-existing baseline-env set (model-selection / memory / secrets / telegram / deadcode-pnpm), none in changed files.Review
🪨 Rune + 🌊 Ronan — the doom-lock-completeness eyes: verify the trigger-split correctly distinguishes ordinary-subagent-returns (reset) from actual continuation-chain delegate-returns (preserve+cap), and the runaway-leash stays intact. Off assembly tip
3084bf16fc; merges clean with PR #991 (#988-sweep — no file overlap).Co-authored-by: Copilot 223556219+Copilot@users.noreply.github.com