Feat/delegation reflex#259
Merged
Merged
Conversation
…arity Two review follow-ups from the conquer falsifier round: (1) unattended --push now refuses to land on a protected branch (main/master by default, protectedPushBranches config override, case-insensitive, fail-closed on an unresolvable HEAD) via isProtectedPushBranch in the forge decision core — a conquer worktree normally sits on its own conquer/<slug> branch, so a protected HEAD means something unexpected happened and the work stays committed locally for a human push. (2) the REPL /conquer handler no longer hardcodes what the CLI exposes: --gate-timeout, --max-hours and --timeout parse through the intent layer and reach the gate spawn and runConquer caps, mirroring the CLI defaults (1800s gate, 600s turn, no wall cap). ⚔️ Forged by [Agon](https://github.com/KERNlang/agon) Co-Authored-By: agon (KERN) <292465531+KERN-Agon@users.noreply.github.com>
Full-roster review (6/6 engines, 5-engine consensus on the blocker):
KERN codegen emits optional array fields as [] into DEFAULT_AGON_CONFIG,
so every real loadConfig() caller passed { protectedPushBranches: [] } and
the Array.isArray fallback never fired — main/master were NOT protected in
production. An empty or whitespace-only list now falls back to the default
['main','master']; only a non-empty override replaces it. Verified against
the real loadConfig() output. Also: CLI --timeout/--gate-timeout reject
negative values instead of propagating them to spawnWithTimeout (agy), the
REPL /conquer rejects --max-turns < 1 exactly like the CLI (minimax), and
the exact production config state is now a test case (zai).
Refuted (no change): the cqAbort cleanup claim — the handler's KERN
cleanup block generates the finally { ctx.setActiveAbort(null) }, and a
still-running background job after /focus-away is the job system's design.
⚔️ Forged by [Agon](https://github.com/KERNlang/agon)
Co-Authored-By: agon (KERN) <292465531+KERN-Agon@users.noreply.github.com>
…vidence RAG roadmap 1b (6/6-engine cesar-mode-rag brainstorm design): when the user's prompt resembles past orchestration runs, the matching episodes (mode, winner, outcome, duration) are injected ahead of the Cesar turn as an [EXPERIENCE] block framed as evidence, never authority. New KERN module cesar/experience.kern: intent summarization (capped, control-char safe), stopword-filtered token-set Jaccard similarity (no sidecar on the hot path), and retrieval gated on a similarity floor AND a min-N support threshold (default 3 qualifying episodes) so a couple of stale runs can't steer routing. recordRun now stores a 160-char intentSummary; legacy records without it are skipped. Injection reuses the shouldGroundInput trigger (slash commands and trivial prompts never retrieve) and is fail-open. Config: cesarExperience (default on), plus tunable minSimilarity/minEpisodes/topK/window resolved 0-safe against the DEFAULT_AGON_CONFIG codegen emission. 14 unit tests. ⚔️ Forged by [Agon](https://github.com/KERNlang/agon) Co-Authored-By: agon (KERN) <292465531+KERN-Agon@users.noreply.github.com>
…path
Full-roster review (6/6 engines) findings, each verified: (1) codex
blocker — telemetry.json is global, so precedent from one repo could
inject its task text into another repo's turn; run records now store a
projectKey (resolved working dir) and retrieval filters on it, fail-closed
for unscoped legacy records. (2) zai/claude/kimi — the new test file
contained raw NUL/ESC bytes and diffed as binary; escaped to \u
sequences. (3) agy — the prompt is tokenized once per retrieval, not once
per episode (scoreTokenSetSimilarity). (4) agy/zai/kimi/claude — the
ledger read is cached on (mtimeMs,size) so unchanged telemetry.json is
parsed once, and append() trims to MAX_LEDGER_RUNS (2000) so the file
cannot grow unbounded. (5) kimi — tokenization is unicode-aware
(\p{L}\p{N}) so non-English prompts score. (6) codex prompt-replay:
summaries are the user's own local prompts (same trust domain); quote
normalization added for rendering hygiene. Also: buildExperienceBlock
rewritten loop-style to satisfy the kern:self-coverage var-bad-expr
ratchet (every blocker category sits exactly at its cap).
⚔️ Forged by [Agon](https://github.com/KERNlang/agon)
Co-Authored-By: agon (KERN) <292465531+KERN-Agon@users.noreply.github.com>
… tasks First slice of the 6/6-engine delegation-reflex brainstorm consensus: a pure veto-first scorer (cesar/delegation-reflex.kern) detects prompts with ≥3 explicit list items and NO sequential-dependency, shared-artifact, or small-task markers, and injects an in-prompt structural note that the work may decompose. Surfacing stays with the existing suggestion flow — Cesar may offer [SUGGEST:agent-team], promptDelegation confirms with the user, nothing ever auto-spawns. False positives cost one advisory line by construction: prose, ordered work, 2-item lists, tiny items, and prompts under 80 chars never trigger. Config cesarDelegationReflex (default on), fail-open injection beside the experience block. 10 unit tests. ⚔️ Forged by [Agon](https://github.com/KERNlang/agon) Co-Authored-By: agon (KERN) <292465531+KERN-Agon@users.noreply.github.com>
Full-roster review (6/6 engines) fixes: (1) kimi — the advisory pointed Cesar at [SUGGEST:agent-team], a marker parseSuggestion/routing do not recognize; the canonical action is team-agent (team- prefix + agent). (2) agy — shared-state vetoes now match plural forms (schemas, lockfiles, migrations). (3) kimi — the first…then sequential veto is windowed to one sentence-ish span so unrelated far-apart mentions no longer over-veto. Recorded as by-design: multiline list items measured by their first line (under-triggering is the stated cost asymmetry) and the shouldGroundInput reuse (generic real-prose predicate, and the reflex carries its own 80-char floor). ⚔️ Forged by [Agon](https://github.com/KERNlang/agon) Co-Authored-By: agon (KERN) <292465531+KERN-Agon@users.noreply.github.com>
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
2 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.
No description provided.