fix: narrow ContinuationSignal union before .task access in debug log#56
Closed
cael-dandelion-cult wants to merge 18 commits intofeature/context-pressure-squashedfrom
Closed
Conversation
…ate reservations Adds the continuation feature: CONTINUE_WORK timer-based self-scheduling, continue_delegate tool for background shard dispatch, context-pressure event injection, chain-hop tracking with generation guards, cost cap enforcement, and delayed delegate reservation accounting. Includes DelayedContinuationReservation store with in-memory backing, admission math using highest allocated hop label, and reservation lifecycle (create → timer fire → take → spawn → persist on acceptance). New files: context-pressure.ts, continuation-runtime.ts, continuation-delegate-store.ts, continue-delegate-tool.ts, and comprehensive test coverage for all paths.
…ification Gates continuationTrigger behind continuation.enabled at both announce sites. Fixes setDelegatePending leak on spawn failure (4 cleanup paths). Aligns context-pressure gate to === true (opt-in). Classifies 'continuation' and 'silent-wake-enrichment' as wake reasons in heartbeat scheduler. Preserves descendant completion findings in non-wake announce path. Restores requesterIsInternalSession usage lost in rebase. Protocol models regenerated. Test coverage for all fix paths.
…servation model Updates the continuation RFC to document sessions_yield as a complementary tool, delayed delegate reservation semantics (in-memory backing, admission via highest allocated hop, reservation lifecycle), and aligns prose with the implementation.
Adds system prompt sections for CONTINUE_WORK, continue_delegate, context pressure, cooperative yield (sessions_yield), and delegated continuation. Includes full/minimal split for continuation instructions and matching test expectations.
Timer-fired doSpawn captured accumulatedChainTokens at scheduling time. If an intervening turn raised continuationChainTokens, the timer callback overwrote the higher value with a stale snapshot, regressing cost accounting. Fix: Math.max(captured, current) on both bracket and tool delegate timer paths. Tokens only go up, never regress. Addresses P2 from upstream Codex review.
The refactor to commands/agent.ts dropped the explicit per-run provider/model override path that was present in agents/agent-command.ts. Gateway agent.run calls with explicit model silently used defaults. Add the normalizeExplicitOverrideInput helpers, allowlist validation, and override application so opts.provider/opts.model are properly applied after session-level overrides. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Non-wake heartbeats (periodic polls) were draining [continuation:wake] events via drainFormattedSystemEvents, causing the continuation chain to stall. The dedicated work-wake run that should consume these events never saw them. Peek and preserve [continuation:wake] events before the drain, then re-enqueue them after, so they survive until the actual continuation-wake heartbeat fires. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When both the subagent's own reply and childCompletionFindings are present, the findings variable only contained one or the other. Now merge them so the parent session receives both the subagent output and descendant completion context in the announce result. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The continue_delegate tool enqueues delegates into a module-level store that agent-runner drains after the turn. When called via gateway tools-invoke (which has no drain path), delegates would leak into the next agent-runner turn for the same session. Add hasDelegateDrainPath option to createOpenClawTools and only include the tool when the caller can drain delegates (pi-tools.ts agent-runner path). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
finalizeRaw() returned only visibleText, missing the case where the stream ends mid-prefix before it can be resolved as silent. Apply the same pendingSilentPrefix flush logic that finalize() already uses. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Previously ALL subagent completion announces were tagged with continuationTrigger: "delegate-return" when continuation was enabled. This caused regular subagent completions to preserve chain state and skip timer cancellation, which is only appropriate for continuation delegates. Now only tag as delegate-return when the completing run was a continuation delegate (task contains chain-hop marker or was a silent announce). Queue-based announces no longer carry the trigger at all. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…th ran bracketAlreadyAccumulated was set based on continuationSignal presence, but the bracket path may not actually accumulate tokens (e.g. when the chain is capped). This caused the tool-delegate path to skip token accumulation incorrectly. Now track whether the bracket path actually accumulated tokens with an explicit flag. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Silent-wake delegates (| silent-wake) wake the parent to act on the enrichment, but the steer instruction told the agent to reply NO_REPLY. This contradicts the purpose of silent-wake. Now generate a distinct instruction for silent-wake enrichment that tells the agent to process and act on the result rather than suppressing all output. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Add provider, model, allowModelOverride fields to commands/agent/types.ts AgentCommandOpts and AgentCommandIngressOpts to match agents/command/types.ts - Fix resolveAgentRunTrigger return type to match EmbeddedRunTrigger - Add allowModelOverride validation to agentCommandFromIngress - Remove unreachable silentAnnounce check in P2-3 fix (TS2367) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
P1-6: Replace hasDelegateDrainPath with drainsContinuationDelegateQueue and thread it through the full call chain (openclaw-tools, pi-tools, params, attempt, get-reply-run, queue types) so only runs that actually drain the staged delegate queue expose continue_delegate. P2-1: Extract flushPendingSilentPrefix() helper inside createAcpVisibleTextAccumulator to DRY the finalize/finalizeRaw logic. Export __testing for unit test access and add agent-command.test.ts. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Logs payload count, last payload shape, text tail, and whether a bracket delegate token exists in a non-last payload. This will diagnose the bracket delegate parsing failure where the token appears in the response but the parser doesn't extract it. Uses continuation/guard subsystem logger at debug level. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The parser checked payloadArray[last].text, but when tool calls interleave, the bracket token ends up in a non-last payload. This breaks subagent chain-hops where the bracket is the ONLY continuation path (continue_delegate tool is denied for subagents). Fix: walk backwards through payloads to find the last one with text. Diagnostic logging preserved (gated on signal detection, debug level). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Fixes TS2339 build error on feature/context-pressure-squashed.
karmafeast
pushed a commit
that referenced
this pull request
Mar 29, 2026
Fixes TS2339: .task only exists on kind:"delegate", not kind:"work". Guard with "task" in signal before accessing. Closes #56 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
karmafeast
pushed a commit
that referenced
this pull request
Mar 30, 2026
Issue #54: continue_delegate tool invisible in prince toolsets. Root cause: drainsContinuationDelegateQueue flag dropped during field-by-field param reconstruction in run.ts, agent-runner-utils.ts, agent-runner-execution.ts, and followup-runner.ts. Fixed all 4 relay sites. Added permanent logVerbose diagnostic at the gate to catch future flag-threading regressions (fires for undefined, not false). Issue #56: TS2339 from .task access on ContinuationSignal union without narrowing. Fixed with signal.kind === "delegate" discriminant guard. Reviewed-By: Elliott 🌻, Ronan 🌊, Silas 🌫️, Cael 🩸 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
4f3461e to
09abc00
Compare
cael-dandelion-cult
added a commit
that referenced
this pull request
Apr 28, 2026
… canonical2 (per figs 14:05 PDT) (#421) * docs: seed release highlights sync doc * docs(release-highlights): add 🌫 #56 cross-walk + config-bits + 8 uncovered TC list * docs(release-highlights): 🌻 commit-delta walk (E1–E10) + cross-cutting flags * docs(release-highlights): 🌊 merge editor pass — canonical case board + (c)/(d) columns Dedup: 🌫 TC-* labels collapsed into 🌻 E6.{1,2,3,4} per 🌫 vote, aliases preserved. Adds: (c) swim-37 case-stub + (d) RFC-appendix slot per highlight. Locks: 32-case canonical board (24 E-series + 5 B-twins + 3 D-cfg + 1 TC-no-genguard). Pins: TC-no-genguard §3.2 not §3.6; phantom cleanup-debt retracted; feedback-key clarification; maxChainLength boundary-pin. Surfaces: 2 figs-pending Qs (X1 echo, C1 chain-to-root). Standing for cohort second-eye: 🌫 / 🌻 / 🩸. * docs(release-highlights): fix case-count math (36 not 32) per 🩸 second-eye * docs(release-highlights): +D-cfg.sdq-retry-not-hot-reloadable per 🌻 second-eye RFC §3.6/§6.5 documents session-delivery-queue.retry.cap + .backoffMs[] but they're not hot-reloadable in the impl. Add as 37th case to catch the docs/code-shape mismatch instead of leaving as silent footnote. * docs(release-highlights): +E6.5 descriptor-content regression per 🌫 second-eye Snapshot test on continue_delegate tool-description JSON catches descriptor-string drift (substrate-naming line, bc#11 cross-link, targetSessionKey listing) independent of runtime delivery path. Lightweight; covers #336/#338 descriptor-content axis. * docs(release-highlights): pin total at 35 net-new cases per cohort convergence E9 + E10 reframed as covered-by-static-harness traceability rows, not net-new swim-37 work. 25 E-series + 5 B + 4 D-cfg + 1 TC = 35. 🌫 + 🌻 + 🩸 cohort acked 35 as the canonical number. * docs(release-highlights): pin total at 36 (E1-E7 = 26 incl E6.5) per 🩸 byte-check Off-by-one in prior 35: E6.x = 5 (incl E6.5 add), not 4. 26 net-new E + 5 B + 4 D-cfg + 1 TC = 36. * docs(release-highlights): pin total at 37 per cohort convergence 🌫 + 🌻 + 🩸 all reconverged on 37 table-rows (regex-trap on E9/E10 cleared). Dropping net-new vs table-rows split to match cohort framing; E9/E10 still footnoted as static-harness covered. * docs(release-highlights): pin total at 38 (E9/E10 + E6.5 all counted) 🌫 byte-verified 28 E + 5 B + 4 D-cfg + 1 TC = 38 on 162bb97. My earlier 37 dropped E6.5 from the E-bucket math; 27→28 with E6.5 included. Final cohort signoff: 🌻 ✅ + 🌫 ✅ + 🩸 ✅ pending final-touch. * docs(release-highlights): pin both numbers per 🌫 break-the-loop proposal board_total: 38 (rows on canonical board) net_new_swim_cases: 36 (38 minus E9/E10 static-harness coverage) Disagreement during 08:04-08:18 PDT convergence was categorical (which bucket E6.5/E9/E10 land in), not arithmetic. Two pinned numbers prevents re-derivation. * docs(release-highlights): +TC-tools-registered-post-deploy (figs 08:54 PDT) figs-flagged smoke: positive presence-of-tools assertion post-deploy on candidate prince. Catches the 'all green with one tool missing' precedent horror. Pairs with TC-continuation-default-off (negative case). board_total: 38 → 39 net_new_swim_cases: 36 → 37 * docs(release-highlights): rename TC-tools-registered → D-tools.continue-delegate-registered Per 🌻 + 🩸 convergence: D-tools is the right bucket (deploy/runtime-surface, not test-case shape). Broadened scope per 🌻's concrete proposal — registry probe asserting all three tools present in manifest. Totals unchanged: board_total: 39 / net_new_swim_cases: 37 * docs(release-highlights): +sessions_yield to D-tools manifest probe (🌫 catch) * docs(release-highlights): explicit base-matrix reference per figs 08:56 PDT figs catch: 'TOTALITY of test cases, weve ASSEMBLED them more than twice... SWIM/ directory... should not be conjuring it fresh every time.' Pin the swim-37 overlay (39 rows) as DELTA on top of the durable base in openclaw-bootstrap/SWIM/FORMAL-SWIM-RUNBOOK.md §4 + #427 + #412. Full swim-37 matrix = base ∪ overlay. Not a substitute for canonical. --------- Co-authored-by: Silas 🌫 <silas.dandelion.cult@hotmail.com> Co-authored-by: elliott-dandelion-cult <elliott.dandelion.cult@gmail.com> Co-authored-by: Ronan 🌊 <ronan@solidor.io>
karmafeast
pushed a commit
that referenced
this pull request
May 1, 2026
… canonical2 (per figs 14:05 PDT) (#421) * docs: seed release highlights sync doc * docs(release-highlights): add 🌫 #56 cross-walk + config-bits + 8 uncovered TC list * docs(release-highlights): 🌻 commit-delta walk (E1–E10) + cross-cutting flags * docs(release-highlights): 🌊 merge editor pass — canonical case board + (c)/(d) columns Dedup: 🌫 TC-* labels collapsed into 🌻 E6.{1,2,3,4} per 🌫 vote, aliases preserved. Adds: (c) swim-37 case-stub + (d) RFC-appendix slot per highlight. Locks: 32-case canonical board (24 E-series + 5 B-twins + 3 D-cfg + 1 TC-no-genguard). Pins: TC-no-genguard §3.2 not §3.6; phantom cleanup-debt retracted; feedback-key clarification; maxChainLength boundary-pin. Surfaces: 2 figs-pending Qs (X1 echo, C1 chain-to-root). Standing for cohort second-eye: 🌫 / 🌻 / 🩸. * docs(release-highlights): fix case-count math (36 not 32) per 🩸 second-eye * docs(release-highlights): +D-cfg.sdq-retry-not-hot-reloadable per 🌻 second-eye RFC §3.6/§6.5 documents session-delivery-queue.retry.cap + .backoffMs[] but they're not hot-reloadable in the impl. Add as 37th case to catch the docs/code-shape mismatch instead of leaving as silent footnote. * docs(release-highlights): +E6.5 descriptor-content regression per 🌫 second-eye Snapshot test on continue_delegate tool-description JSON catches descriptor-string drift (substrate-naming line, bc#11 cross-link, targetSessionKey listing) independent of runtime delivery path. Lightweight; covers #336/#338 descriptor-content axis. * docs(release-highlights): pin total at 35 net-new cases per cohort convergence E9 + E10 reframed as covered-by-static-harness traceability rows, not net-new swim-37 work. 25 E-series + 5 B + 4 D-cfg + 1 TC = 35. 🌫 + 🌻 + 🩸 cohort acked 35 as the canonical number. * docs(release-highlights): pin total at 36 (E1-E7 = 26 incl E6.5) per 🩸 byte-check Off-by-one in prior 35: E6.x = 5 (incl E6.5 add), not 4. 26 net-new E + 5 B + 4 D-cfg + 1 TC = 36. * docs(release-highlights): pin total at 37 per cohort convergence 🌫 + 🌻 + 🩸 all reconverged on 37 table-rows (regex-trap on E9/E10 cleared). Dropping net-new vs table-rows split to match cohort framing; E9/E10 still footnoted as static-harness covered. * docs(release-highlights): pin total at 38 (E9/E10 + E6.5 all counted) 🌫 byte-verified 28 E + 5 B + 4 D-cfg + 1 TC = 38 on 162bb97. My earlier 37 dropped E6.5 from the E-bucket math; 27→28 with E6.5 included. Final cohort signoff: 🌻 ✅ + 🌫 ✅ + 🩸 ✅ pending final-touch. * docs(release-highlights): pin both numbers per 🌫 break-the-loop proposal board_total: 38 (rows on canonical board) net_new_swim_cases: 36 (38 minus E9/E10 static-harness coverage) Disagreement during 08:04-08:18 PDT convergence was categorical (which bucket E6.5/E9/E10 land in), not arithmetic. Two pinned numbers prevents re-derivation. * docs(release-highlights): +TC-tools-registered-post-deploy (figs 08:54 PDT) figs-flagged smoke: positive presence-of-tools assertion post-deploy on candidate prince. Catches the 'all green with one tool missing' precedent horror. Pairs with TC-continuation-default-off (negative case). board_total: 38 → 39 net_new_swim_cases: 36 → 37 * docs(release-highlights): rename TC-tools-registered → D-tools.continue-delegate-registered Per 🌻 + 🩸 convergence: D-tools is the right bucket (deploy/runtime-surface, not test-case shape). Broadened scope per 🌻's concrete proposal — registry probe asserting all three tools present in manifest. Totals unchanged: board_total: 39 / net_new_swim_cases: 37 * docs(release-highlights): +sessions_yield to D-tools manifest probe (🌫 catch) * docs(release-highlights): explicit base-matrix reference per figs 08:56 PDT figs catch: 'TOTALITY of test cases, weve ASSEMBLED them more than twice... SWIM/ directory... should not be conjuring it fresh every time.' Pin the swim-37 overlay (39 rows) as DELTA on top of the durable base in openclaw-bootstrap/SWIM/FORMAL-SWIM-RUNBOOK.md §4 + #427 + #412. Full swim-37 matrix = base ∪ overlay. Not a substitute for canonical. --------- Co-authored-by: Silas 🌫 <silas.dandelion.cult@hotmail.com> Co-authored-by: elliott-dandelion-cult <elliott.dandelion.cult@gmail.com> Co-authored-by: Ronan 🌊 <ronan@solidor.io>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Fixes TS2339 build error introduced by
3d452ad/7ad39ef(codex session 2026-03-28).Line 857 of
agent-runner.tsaccesses.taskon the fullContinuationSignalunion without narrowing tokind: "delegate". Theworkvariant doesn't have.task.Fix:
"task" in continuationResult.signalguard before access.Build passes clean after this fix.
Ref: #54