From edd30d7194251ebedff2f9a03ecda7b116744584 Mon Sep 17 00:00:00 2001
From: ParthSareen <parth.sareen@ollama.com>
Date: Thu, 11 Jun 2026 21:03:47 -0700
Subject: [PATCH] docs: add agent improvements report

---
 improvements-report.md | 109 +++++++++++++++++++++++++++++++++++++++++
 1 file changed, 109 insertions(+)
 create mode 100644 improvements-report.md

diff --git a/improvements-report.md b/improvements-report.md
new file mode 100644
index 000000000..2253c1a66
--- /dev/null
+++ b/improvements-report.md
@@ -0,0 +1,109 @@
+# Agent Loop Improvements Report
+
+Base: `ad5dccd3f` on `parth-agent-loop`.
+
+Verification used after each task:
+
+- `go build ./...` passed
+- `go test -count=1 ./agent/... ./cmd/... ./api/...` passed
+
+## Task 1: Unify approval dispatch
+
+- Added `RequiresApproval(context.Context, Tool, ApprovalRequest) bool` to `agent.ApprovalHandler`.
+- Removed the session concrete-type leak for `*ApprovalManager` and the private approval requirement interface.
+- Added a wrapper-handler regression test proving policy is consulted through the public handler contract.
+
+Commit: `575845472 agent: unify approval dispatch`
+
+## Task 2: Persist images; drop dead schema columns
+
+- Added `messages.images` storage and migration for existing SQLite chat DBs.
+- Serialized/deserialized `api.Message.Images` on append, update, and reload.
+- Removed dead `model_cloud` / `model_ollama_host` columns from new table creation.
+
+Commit: `5272b3fd8 agent: persist chat images`
+
+## Task 3: Delete orphaned REPL
+
+- Removed the unreachable legacy interactive REPL path and dead raw terminal chat/generate helpers.
+- Kept the still-used multimodal file extraction helpers in `cmd/file_data.go`.
+- Removed tests for the deleted create-request save path.
+
+Commit: `6384fd304 cmd: remove orphaned repl`
+
+## Task 4: Approval prompt over modals
+
+- Approval prompts now close resume/history modals before opening.
+- Added a regression test for approval arriving while history is open.
+
+Commit: `9f90698ab tui: surface approvals over modals`
+
+## Task 5: Live working directory state
+
+- Added `chatModel.workingDir` for live cwd display/session state.
+- Removed event-path writes to `m.opts.WorkingDir`.
+- Added a test proving tool-finished cwd events update the footer without mutating construction-time options.
+
+Deliberately not fixed: `chatModel` still mutates a few non-cwd option fields for existing flows like resume, skills, permission mode, and context-window refresh. That broader separation is outside this cwd-specific task.
+
+Commit: `32f030eaf tui: track live working directory`
+
+## Task 6: Batch streaming persistence
+
+- Replaced per-delta assistant persistence with thresholded flushes plus forced final flush on success/cancel.
+- Added `idx_messages_chat_id_id` for last-message update lookup.
+- Added tests for bounded write count and partial canceled-stream persistence.
+
+Commit: `c61134c53 agent: batch streaming persistence`
+
+## Task 7: Merge model-aware chat store
+
+- Collapsed `ModelAwareChatStore` into `ChatStore`.
+- Renamed SQLite store methods to `AppendMessage(..., model)` / `UpdateLastMessage(..., model)`.
+- Removed session type assertions and thin no-model wrappers.
+
+Commit: `27d3e78ff agent: merge model-aware chat store`
+
+## Task 8: Decompose chat TUI
+
+- Split `cmd/tui/chat.go` into focused same-package files for approval, events, input, markdown, modals, and rendering.
+- Split `chat_test.go` along matching seams while leaving shared test fakes/helpers in the base file.
+- No intended behavior changes.
+
+Commit: `958271d75 tui: split chat implementation`
+
+## Task 9: Harden bash approval classification
+
+- Added explicit threat-model comment: static command analysis, not sandboxing.
+- Classified function declarations, `eval`, `source`/`.`, `exec`, and dynamic command names as high risk.
+- Added tests for shell-function, eval, variable-command, and command-substitution evasions.
+
+Commit: `330c8bbe3 agent: harden bash approval classification`
+
+## Task 10: Web tools require approval
+
+- `web_search` and `web_fetch` now self-declare approval.
+- Default policy prompts low-risk per query/URL with target-specific session keys.
+- Added policy and tool declaration tests.
+
+Commit: `a749038dd agent: require approval for web tools`
+
+## Task 11: Validate bash returned cwd
+
+- `readFinalWorkingDir` now returns a cwd only if the captured path exists and is a directory.
+- Added tests for regular-file, missing, and valid directory paths.
+
+Commit: `6d80dfd94 agent: validate bash working directory`
+
+## Task 12: Bound caches and summaries
+
+- Replaced unbounded markdown renderer `sync.Map` with an 8-entry width cache.
+- Truncated compaction summaries over 16KB with a marker before storing/returning.
+- Added cache-size and oversized-summary tests.
+
+Commit: `78fb5c426 agent: bound markdown and compaction caches`
+
+## Final Notes
+
+- Final verification passed with the same build/test gate.
+- Linker warnings about duplicate native libraries appeared during builds, but builds and tests completed successfully.