fix(cesar): remove the no-project-brief warning nudge - #291
Merged
Conversation
The once-per-session 'No project brief found' warning was noisy and its message text was stale — the brief cascade accepts AGENTS.md and CLAUDE.md (not just AGENTS.md/.agon/project.md as the message claimed), and the cwd-only check misfired in subdirectories. Removed the nudge block, the _noBriefNudged WeakMap, and the now-unused hasProjectBrief import; the core hasProjectBrief export and the verify-before-done gate subsystem are untouched. ⚔️ Forged by [Agon](https://github.com/KERNlang/agon) Co-Authored-By: agon (KERN) <292465531+KERN-Agon@users.noreply.github.com>
The one-time 'agon now runs engines in workspace-pure mode' startup banner claimed dispatched engines no longer get your CLAUDE.md — stale since the project-context scanner injects AGENTS.md/CLAUDE.md into every dispatch, and noisy when the config latch failed to persist. Removed the banner, its now-dead isolationMigrationNotified config field (loadConfig tolerates the stale key in existing configs), and the TUI-probe fixture lines. The workspace-pure default and its --impure / engineIsolation escape hatch are now documented durably in the README instead of a runtime nag. ⚔️ Forged by [Agon](https://github.com/KERNlang/agon) Co-Authored-By: agon (KERN) <292465531+KERN-Agon@users.noreply.github.com>
…tive engines API/native engines are the normal Cesar brains, but the native tool loop had no notion of earned progress and no mid-loop steering channel: a read-only mapping spiral ran 29/30 steps for 31 minutes (Bash cat/grep bypassed the Read/Grep/Glob-only search counter, every tool event counted as progress, budget grew on 'didn't fail'), and steering typed mid-turn sat undelivered until Esc. Now every step is classified by effect (read/verify/mutate/other; Bash via isReadOnlyCommand + discovered-gate matchers), with one canonical step signature shared by all ledgers: - Bash reads join the searchToolCount ledger, so the CODEBASE-BRIEF advisory fires for shell-readers. - Budget growth is earned only by mutate/verify/novel-read/other steps; read-repeats never grow the budget (spirals stall at base 15). - Intent-aware one-shot read-spiral note: edit turns get 'summarize, then implement or ask'; investigate turns are never told to implement and only trip on read-repeat volume. - No-progress signature counts novelty (novelStepCount) plus a monotonic substantive-answer latch instead of raw tool events, so re-read rounds strike toward the existing checkpoint while real mapping and real answers count as progress. - Steering: session-resume gains an optional shouldYield hook polled at every completed tool-pair boundary; pending steering ends the cycle cleanly (done reason steering-yield, in-flight calls never aborted), is delivered as a real user turn with its images, refunds the continuation slot, and never strikes no-progress. - Thresholds are config policy (cesarSearchNudgeThreshold / cesarReadSpiralThreshold / cesarReadRepeatThreshold), defaults 40/25/12. Design ran the full ladder: nero killed the count-based v1 (FLAWED), a 6-engine brainstorm converged on effect classification + earned budget with steering as a genuine user turn, and the 6-engine role review's 10 findings were verified — 8 fixed with discriminating tests, 2 dismissed with evidence. Spec: .claude/specs/cesar-native-loop-guards/spec.md ⚔️ Forged by [Agon](https://github.com/KERNlang/agon) Co-Authored-By: agon (KERN) <292465531+KERN-Agon@users.noreply.github.com>
Four verified findings from the pre-push branch review, all fixed with
discriminating tests (+17, now 4732):
- Continuation sends (steering delivery, auto-continue, closure,
confirmation) ignored native tool_call chunks, so steered writes
executed without mutation/verification/progress/display tracking and
could dodge the verify-before-done gate. One shared
_forwardContinuationToolCall helper now records them at all 7 sites,
gated on hasNativeTools to avoid double-counting on XML engines.
- Read step signatures now carry offset/limit (and Grep/Glob their
scope args), so paging a large file earns budget instead of counting
as read-repeats.
- A zero-text steering-yield on the initial send now always reaches the
delivery loop instead of waiting for a later turn.
- Successful unclassifiable ('other') Bash steps now suppress the
edit-intent read-spiral note via a separate shellWorkStepCount —
shell work we can't prove read-only must not draw a "you haven't
implemented" nudge; effectfulStepCount keeps its strict
mutate/verify meaning for gate logic.
Dismissed with evidence: the reported block-scope ReferenceError in
the steering drain (typecheck-proven in-scope), the "gate matchers
never assigned" claim (assigned at turn start, read lazily), and the
stale-signature strike after a yield refund (all signature components
are monotonic — a stale baseline can only read as progress).
⚔️ Forged by [Agon](https://github.com/KERNlang/agon)
Co-Authored-By: agon (KERN) <292465531+KERN-Agon@users.noreply.github.com>
The plan-approval selector rendered ABOVE the plan markdown in the pinned lower panel, so a multi-screen plan buried the question the user had to answer. Now the plan body is committed as a normal scrollable transcript block at proposal time and a compact 5-row PlanApprovalPrompt is pinned directly above the composer; the bottom chrome row budget accounts for exactly those 5 rows, and tall plans seal into scrollback instead of inflating the fixed region. Review of the change surfaced that rejecting a Cesar proposal was entirely broken: both the keyboard '3. Reject' and /cancel routed to the legacy plan runner's handleCancel, which reads ctx.currentPlan, answered 'No active plan to cancel.' and left the proposal pinned. New cancelPendingCesarPlan mirrors the approve path: cancels + saves the plan, rewrites the on-disk record, releases the prompt, and appends a [plan rejected] note to the chat session so Cesar can't silently re-propose. plan-dismiss now carries an outcome and the transcript appends a one-line record for superseded plans, so the committed 'Status: awaiting_approval' snapshot can no longer lie about the final state. ⚔️ 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.
The once-per-session 'No project brief found' warning was noisy and its message text was stale — the brief cascade accepts AGENTS.md and CLAUDE.md (not just AGENTS.md/.agon/project.md as the message claimed), and the cwd-only check misfired in subdirectories. Removed the nudge block, the _noBriefNudged WeakMap, and the now-unused hasProjectBrief import; the core hasProjectBrief export and the verify-before-done gate subsystem are untouched.
⚔️ Forged by Agon