feat(agent-loop): snip-tokens-freed feedback loop (#4) - #222
Merged
Conversation
The pre-send snip (cap_oversized_tool_results) freed tokens but never reported how many, so the post-response fold always fired at 75% even when the snip had just bought plenty of headroom. - Add `cap_oversized_tool_results_counted` — wraps the unchanged capper and reports tokens freed (measured with the same estimator the fold decision uses). Existing callers stay on the Vec-returning fn. - run_loop tracks `snip_tokens_freed` from the (now tiered) cap site. - In the post-usage Fold path, a pure `snip_bought_enough(freed, ctx_max, aggressive)` skips a NORMAL fold when the snip freed > SNIP_SUFFICIENT_FRACTION (10%) of the window. Aggressive / force-summary folds still fire. The credit resets after each post-usage decision so a stale snip can't suppress a later fold. Tests: snip_bought_enough gating (normal vs aggressive, <10%, div-0 guard) and cap_counted freed-token accuracy. 2138 pass at -D warnings. Stacked on the aggressive-prune branch (PR #221).
This was referenced May 29, 2026
yogthos
changed the base branch from
feat/improve-2-aggressive-prune-tier
to
main
May 29, 2026 21:10
allen-munsch
pushed a commit
to allen-munsch/dirge
that referenced
this pull request
May 29, 2026
…VEMENTS_PLAN #2) After a summary fold replaced the middle of the conversation, the model lost the concrete contents of the files it was editing and had to re-read them (or acted on stale assumptions). - FileTouchTracker::working_files() exposes the tracked overlap set (sorted). - SummaryOutcome::Succeeded now carries the summary insertion index. - compression::build_post_compact_snapshots (pure) builds `[Post-compaction file snapshot: <path>]` system messages, capped at POST_COMPACT_MAX_FILES (5) and POST_COMPACT_MAX_TOKENS_PER_FILE (5000, head+tail truncated) so restoration can't itself trigger a fold. - run_loop::restore_working_files re-reads those files (tokio::fs, thin I/O, unreadable files skipped) and splices the snapshots right after the summary message, at all three fold sites. Chose a pure builder + thin in-loop read over a LoopConfig reader closure: same testability for the meaningful logic (cap + truncation + message shape) without the Clone/Debug/Default/builder surface a new config field would add. Tests: working_files() sorted overlap; snapshot builder caps count + truncates oversized files. 2140 pass at -D warnings. Stacked on the snip-feedback branch (PR dirge-code#222).
yogthos
added a commit
that referenced
this pull request
Jun 2, 2026
(#358) Themes can now set a 'background' color, painted behind every cell while preserving foregrounds (selection's REVERSED modifier still swaps correctly). phosphor defaults to a near-black charcoal #222222 (CRT-bezel depth); plain and the opt-out use Color::Reset (keep the terminal's own background, no fill). Wiring: Theme.background field + both presets + ThemeJson key + merge_into + theme::background() accessor; Scene carries it; render_frame patches the whole area with Style::default().bg(..) after all widgets (ratatui patches only the bg, so fg + REVERSED are untouched), skipped entirely when Reset. + scene test (fills when set / no-op on Reset) + docs/themes.md row. Co-authored-by: Yogthos <yogthos@gmail.com>
yogthos
pushed a commit
that referenced
this pull request
Jun 3, 2026
Adds DAP step-through debugging, the phased /plan workflow, minified read/edit, configurable theme background (#222 phosphor default), and snap-to-bottom; plus a round of UI bug fixes (resize scroll clamp, idle Ctrl+C draft, Home/End rebind, /plan busy strand).
allen-munsch
pushed a commit
to allen-munsch/dirge
that referenced
this pull request
Jun 3, 2026
…-feedback-loop feat(agent-loop): snip-tokens-freed feedback loop (#4)
allen-munsch
pushed a commit
to allen-munsch/dirge
that referenced
this pull request
Jun 3, 2026
…VEMENTS_PLAN #2) After a summary fold replaced the middle of the conversation, the model lost the concrete contents of the files it was editing and had to re-read them (or acted on stale assumptions). - FileTouchTracker::working_files() exposes the tracked overlap set (sorted). - SummaryOutcome::Succeeded now carries the summary insertion index. - compression::build_post_compact_snapshots (pure) builds `[Post-compaction file snapshot: <path>]` system messages, capped at POST_COMPACT_MAX_FILES (5) and POST_COMPACT_MAX_TOKENS_PER_FILE (5000, head+tail truncated) so restoration can't itself trigger a fold. - run_loop::restore_working_files re-reads those files (tokio::fs, thin I/O, unreadable files skipped) and splices the snapshots right after the summary message, at all three fold sites. Chose a pure builder + thin in-loop read over a LoopConfig reader closure: same testability for the meaningful logic (cap + truncation + message shape) without the Clone/Debug/Default/builder surface a new config field would add. Tests: working_files() sorted overlap; snapshot builder caps count + truncates oversized files. 2140 pass at -D warnings. Stacked on the snip-feedback branch (PR dirge-code#222).
allen-munsch
pushed a commit
to allen-munsch/dirge
that referenced
this pull request
Jun 3, 2026
…irge-code#222 (dirge-code#358) Themes can now set a 'background' color, painted behind every cell while preserving foregrounds (selection's REVERSED modifier still swaps correctly). phosphor defaults to a near-black charcoal #222222 (CRT-bezel depth); plain and the opt-out use Color::Reset (keep the terminal's own background, no fill). Wiring: Theme.background field + both presets + ThemeJson key + merge_into + theme::background() accessor; Scene carries it; render_frame patches the whole area with Style::default().bg(..) after all widgets (ratatui patches only the bg, so fg + REVERSED are untouched), skipped entirely when Reset. + scene test (fills when set / no-op on Reset) + docs/themes.md row. Co-authored-by: Yogthos <yogthos@gmail.com>
yogthos
added a commit
that referenced
this pull request
Jun 3, 2026
#371) The file-completion and rewind pickers wrote raw cursor-positioned escapes to std::io::stdout() — but TerminalGuard dup2's fd 1 to the log for the whole TUI while ratatui paints via a separate /dev/tty fd. So the picker candidate-LIST overlay never reached the screen (it went to the log); only the typed @query and the inserted selection were visible via the input buffer. Investigation confirmed the fd-isolation commit (83e3ea9) orphaned the pickers without giving them a ratatui rendering path. Render the overlay through the scene instead: pickers now expose an owned PickerOverlay snapshot (title/rows/selected/empty_hint); the renderer caches it (draw_bottom syncs the file picker from the input editor automatically, the rewind picker is set/cleared explicitly) and paint_picker_overlay draws the windowed list over the bottom rows of the chat region just above the input box. It's diff-safe (ratatui owns those cells) and inherits the theme #222 background fill — fixing both the invisibility and the original wrong-bg report. Removes ~150 lines of direct-stdout draw code and the now-dead input_top_row/color helpers. Adds 2 render tests (candidate list + selected marker; empty 'no matches' hint). 2471 default + 2565 all-features tests pass; clean under -D warnings. Co-authored-by: Yogthos <yogthos@gmail.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
1 participant
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.
Phase 3 of IMPROVEMENTS_PLAN. Stacked on #221 (base = aggressive-prune branch).
Problem
The pre-send snip freed tokens silently; the 75% post-response fold fired regardless of whether the snip had already bought enough headroom — wasting a summarizer call.
Fix
cap_oversized_tool_results_countedwraps the unchanged capper and reports freed tokens (measured withestimate_messages_tokens, the same estimator the fold decision uses). No churn to the ~12 existing callers.run_looptrackssnip_tokens_freedfrom the tiered cap site.snip_bought_enough(freed, ctx_max, aggressive)skips a normal fold when the snip freed >SNIP_SUFFICIENT_FRACTION(10%) of the window. Aggressive / force-summary folds always proceed. Credit resets after each post-usage decision (no stale carry-over).Tests: gating logic (normal vs aggressive, <10%, div-0) + freed-token accuracy. 2138 pass at
-D warnings.