feat: per-agent persistent memory (#62) - #63
Merged
Conversation
Add memory field to AgentConfig that points to a markdown file. On tt up, claude-code agents get memory appended to CLAUDE.md in their worktree; other runtimes get memory prepended to their prompt. The memory directory (.tutti/memory/) is created by ensure_tutti_dir. Memory paths are validated to be relative and non-traversing. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
|
Note Reviews pausedIt looks like this branch is under active development. To avoid overwhelming you with review comments due to an influx of new commits, CodeRabbit has automatically paused this review. You can configure this behavior by changing the Use the following commands to manage reviews:
Use the checkboxes below for quick actions:
📝 WalkthroughWalkthroughAdds an optional per-agent persistent memory field ( Changes
Sequence Diagram(s)sequenceDiagram
participant CLI as CLI (tt up)
participant Config as Config Loader
participant FS as File System
participant Runtime as Agent Runtime
CLI->>Config: load agent configs (may include memory path)
Config-->>CLI: return AgentConfig
alt agent has memory path
CLI->>FS: validate path (relative, no "..", not symlink)
FS-->>CLI: validation result
CLI->>FS: read memory file
FS-->>CLI: memory contents
alt runtime == claude-code
CLI->>FS: inject MEMORY_SECTION into CLAUDE.md (idempotent)
FS-->>CLI: CLAUDE.md updated
CLI-->>Runtime: launch runtime (uses injected CLAUDE.md)
else other runtime
CLI->>CLI: prepend memory to prompt → effective_prompt
CLI-->>Runtime: launch with effective_prompt
end
else no memory
CLI-->>Runtime: launch with original prompt
end
Runtime-->>CLI: runtime started / status (warnings if Unknown/auth issues)
Estimated code review effort🎯 4 (Complex) | ⏱️ ~45 minutes Poem
🚥 Pre-merge checks | ✅ 3 | ❌ 2❌ Failed checks (1 warning, 1 inconclusive)
✅ Passed checks (3 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches🧪 Generate unit tests (beta)
📝 Coding Plan
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 3
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@src/cli/up.rs`:
- Around line 173-178: run_all() currently bypasses the memory injection and
prompt-prepend logic present in run(); update run_all() to call
inject_agent_memory(project_root, &working_dir, agent, &runtime_name)? before
launching each agent and use prepend_memory_to_prompt(project_root, agent,
&runtime_name, agent.prompt.as_deref())? to build the effective_prompt (same as
run()), replacing direct uses of agent.prompt when spawning agents; ensure you
reference inject_agent_memory and prepend_memory_to_prompt and handle their
Result returns the same way run() does.
- Around line 447-452: The inject_agent_memory function currently always appends
to working_dir/CLAUDE.md which can mutate the workspace root and duplicate
entries for reused worktrees; update inject_agent_memory to first detect whether
working_dir is a real worktree path (e.g., compare to project_root or check for
a worktree marker) and only proceed when it’s an actual worktree; write to an
agent-scoped file (include agent.id or AgentConfig name instead of unconditional
CLAUDE.md in the project root) and implement logic to find and replace or update
a managed memory section (bounded markers) inside that file instead of always
appending to avoid duplicate memory on relaunch (apply the same fix to the other
append block referenced near the second occurrence).
In `@src/config/mod.rs`:
- Around line 571-594: The current check uses trimmed.contains("..") which
falsely rejects filenames like "notes..md"; instead, parse the path and reject
only when it contains a ParentDir component: in the loop over self.agents
(checking agent.memory and the trimmed variable) replace the contains("..")
check with inspecting Path::new(trimmed).components() and testing for
Component::ParentDir (use std::path::Component), and keep returning
TuttiError::ConfigValidation with the same message when such a component is
found.
ℹ️ Review info
⚙️ Run configuration
Configuration used: Repository UI
Review profile: CHILL
Plan: Pro
Run ID: ace596fe-8984-4585-9e7b-4f87eef99513
📒 Files selected for processing (9)
src/automation/mod.rssrc/budget/mod.rssrc/cli/doctor.rssrc/cli/handoff.rssrc/cli/up.rssrc/cli/watch.rssrc/config/defaults.rssrc/config/mod.rssrc/state/mod.rs
send_text previously split text into lines and sent each with a separate Enter keypress. This caused multi-line prompts to be interpreted as individual shell commands when the target was a non-TUI runtime (e.g., codex that had exited to zsh). Fix: paste the entire text as a single tmux buffer with bracketed paste mode (-p flag), then send one Enter. Also add a pre-send status check in tt send to warn when the runtime appears to have exited. Discovered during dogfood run of issue #62 orchestration. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
There was a problem hiding this comment.
Actionable comments posted: 1
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@src/cli/send.rs`:
- Around line 76-88: The pre-send health check currently only warns on
AgentStatus::Unknown; update the logic around runtime::get_adapter /
adapter.detect_status to also treat AgentStatus::AuthFailed(_) as a
non-fatal-but-actionable warning when not launched_via_auto_up: match
AuthFailed(err) alongside Unknown and eprintln a clear, user-facing message
referencing agent_name and the auth error (err) with guidance (e.g.,
re-authenticate or check runtime credentials), so users know the runtime is
unusable and how to fix it.
ℹ️ Review info
⚙️ Run configuration
Configuration used: Repository UI
Review profile: CHILL
Plan: Pro
Run ID: aea7eac2-2018-4f84-8379-ef6b29573d35
📒 Files selected for processing (3)
CLAUDE.mdsrc/cli/send.rssrc/session/tmux.rs
✅ Files skipped from review due to trivial changes (1)
- CLAUDE.md
1. Wire memory injection through run_all() so tt up --all injects memory
2. Make inject_agent_memory idempotent with bounded markers — no
duplicate memory sections on relaunch of reused worktrees
3. Skip CLAUDE.md mutation when working_dir is the project root
(no worktree) to avoid modifying the workspace's own file
4. Use Component::ParentDir for path traversal detection instead of
contains("..") which falsely rejected filenames like notes..md
5. Handle AuthFailed status in pre-send health check (warn user)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
There was a problem hiding this comment.
Actionable comments posted: 2
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@src/cli/up.rs`:
- Around line 471-475: When prepare_agent_working_dir detects working_dir ==
project_root and currently returns early (skipping CLAUDE.md injection), change
that branch so Claude agents don't lose memory: for agents whose engine/id
equals "claude-code" call prepend_memory_to_prompt(...) to add the memory into
the prompt (instead of skipping), or alternatively emit a clear warning/return
Err if injection cannot occur; update the same logic at the other occurrence
(the block around lines 530-533) so both code paths treat "claude-code"
consistently rather than silently returning Ok().
- Around line 478-513: Reject symlinked source and destination paths by
canonicalizing and validating them before any file IO in inject_agent_memory and
prepend_memory_to_prompt: for the source (resolved =
project_root.join(memory_path)), use std::fs::canonicalize and ensure the
canonical path starts with the canonical project_root (or otherwise refuse) and
also reject if the original metadata shows a symlink (symlink_metadata +
file_type().is_symlink()); for the destination (claude_md_path =
working.join("CLAUDE.md")), refuse if the destination itself is a symlink or if
canonicalizing the destination would place it outside the canonical working dir
(and similarly check parent path canonicalization), returning an error instead
of reading or writing when validation fails. Ensure these checks reference
memory_path, project_root, resolved, claude_md_path, inject_agent_memory, and
prepend_memory_to_prompt.
ℹ️ Review info
⚙️ Run configuration
Configuration used: Repository UI
Review profile: CHILL
Plan: Pro
Run ID: 08d183e3-bd1e-4499-a871-823ba456ca5d
📒 Files selected for processing (3)
src/cli/send.rssrc/cli/up.rssrc/config/mod.rs
🚧 Files skipped from review as they are similar to previous changes (1)
- src/cli/send.rs
…le Enter Bracketed paste mode (-p) caused codex TUI to consume the first Enter as "close paste bracket" rather than "submit prompt", leaving the text in the input field without processing. Fix: only use -p for multi-line text (where it's needed to prevent line-by-line shell interpretation), and send a second Enter after bracketed pastes to trigger submission. Single-line commands (unset, export, shell commands) skip bracketed paste entirely to avoid the double-Enter issue. Validated via tt send to codex with multi-line prompt — codex immediately started working without needing manual Enter. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
1. Claude agents no longer lose memory when worktrees are disabled: inject_agent_memory returns bool indicating whether file injection succeeded; prepend_memory_to_prompt falls back to prompt injection for claude-code when file_injected=false 2. Reject symlinked memory source and CLAUDE.md destination paths: validate_no_symlink checks symlink_metadata and canonicalize to prevent arbitrary file reads (symlinked source) and writes (symlinked CLAUDE.md) outside the project root Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
There was a problem hiding this comment.
Actionable comments posted: 1
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@src/cli/up.rs`:
- Around line 478-514: The code currently reads and writes files using resolved
(from project_root.join(memory_path)) and claude_md_path
(working.join("CLAUDE.md")) without validating symlinks or ensuring they point
inside the intended directories; update the logic to canonicalize and validate
both paths (use std::fs::canonicalize or equivalent) before any read/write:
canonicalize resolved and ensure it has project_root.canonicalize() as a prefix,
and canonicalize claude_md_path and ensure it has working.canonicalize() as a
prefix (reject or error if canonicalization fails or the paths escape their base
directories), then proceed to read memory_contents and write combined only after
these checks, keeping the MEMORY_SECTION_START / MEMORY_SECTION_END handling
unchanged.
ℹ️ Review info
⚙️ Run configuration
Configuration used: Repository UI
Review profile: CHILL
Plan: Pro
Run ID: 44d59287-0956-4dc9-afec-6a60db89f737
📒 Files selected for processing (7)
src/cli/up.rssrc/config/defaults.rssrc/config/mod.rssrc/session/tmux.rssrc/state/mod.rssrc/worktree/mod.rstests/send_auto_up_integration.rs
🚧 Files skipped from review as they are similar to previous changes (1)
- src/state/mod.rs
Contributor
Author
|
@coderabbitai review All 7 findings have been addressed across commits ea1e8ae, eee7cdf, and 6301cef:
|
|
🧠 Learnings used✅ Actions performedFull review triggered. |
nutt-adam
dismissed stale reviews from coderabbitai[bot], coderabbitai[bot], coderabbitai[bot], and coderabbitai[bot]
March 17, 2026 20:45
All findings addressed.
Contributor
Author
|
@coderabbitai review |
✅ Actions performedReview triggered.
|
There was a problem hiding this comment.
Actionable comments posted: 2
♻️ Duplicate comments (1)
src/cli/up.rs (1)
522-525:⚠️ Potential issue | 🔴 CriticalNew
CLAUDE.mdwrites still bypass the symlink guard.Line 523 only validates the destination when
CLAUDE.mdalready exists. On the first launch, Line 551 creates the file without ever checking the path, so a symlinkedworking_diror parent directory can still redirect this write outside the intended agent tree. Please validate the destination path beforefs::write, not just the already-existing file.Also applies to: 551-551
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/cli/up.rs` around lines 522 - 525, The symlink guard only runs when claude_md_path.exists(); to fix, call validate_no_symlink(&claude_md_path, "CLAUDE.md", working) (or a helper that verifies all path components) before any fs::write that creates CLAUDE.md so the destination is validated on first launch as well; update the code paths around claude_md_path and the fs::write that creates the file to invoke validate_no_symlink (or a new validate_path_parents helper) prior to writing.
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@src/cli/up.rs`:
- Around line 457-468: Update the ConfigValidation error messages in the
validation branch that checks symlinks and canonical path ownership (the block
using path, allowed_root, canonical, root and returning
TuttiError::ConfigValidation) to include the allowed location and an actionable
remediation hint; for the symlink case (where path.is_symlink()) include
guidance to remove or replace the symlink with a real file under the
workspace/worktree and show the offending path, and for the canonicalization
case include the canonicalized allowed_root value and a short instruction to
move the file into that worktree/allowed location or update configuration so the
path is under the shown allowed_root.
- Around line 1492-1499: Don’t silently drop memory by using unwrap_or_else;
instead propagate the error from prepend_memory_to_prompt so failures are
surfaced (and handled by run()) — replace the unwrap_or_else call that assigns
effective_prompt with a fallible handling (use the ? operator or map_err to add
context) so prepend_memory_to_prompt(...) returns an Err up the stack rather
than quietly returning agent.prompt.clone(); reference prepend_memory_to_prompt,
effective_prompt and run() when making the change.
---
Duplicate comments:
In `@src/cli/up.rs`:
- Around line 522-525: The symlink guard only runs when claude_md_path.exists();
to fix, call validate_no_symlink(&claude_md_path, "CLAUDE.md", working) (or a
helper that verifies all path components) before any fs::write that creates
CLAUDE.md so the destination is validated on first launch as well; update the
code paths around claude_md_path and the fs::write that creates the file to
invoke validate_no_symlink (or a new validate_path_parents helper) prior to
writing.
… run_all
- Include allowed directory and remediation hints in symlink/path validation
errors per project convention ("User-facing errors should include actionable
guidance")
- Replace silent unwrap_or_else in run_all() with explicit error handling that
warns and skips the agent, matching run() behavior
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…efore first write - run_all() now propagates prepend_memory_to_prompt errors with ? (matching run() behavior) instead of silently continuing - Validate CLAUDE.md destination and its parent directory for symlinks before any write, not just when the file already exists Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Contributor
Author
|
Addressed all CodeRabbit findings from the latest review:
5723b82 (prior commit):
@coderabbitai review |
|
✅ Actions performedReview triggered.
|
There was a problem hiding this comment.
Actionable comments posted: 1
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@src/cli/up.rs`:
- Around line 539-543: The current removal logic for the managed section uses
existing.find(MEMORY_SECTION_END) which searches from the file start and can
match an unrelated end marker; change it to search for MEMORY_SECTION_END
relative to the found start (e.g. search within &existing[start..] and compute
the absolute end index by adding the start offset) so that the slice for
before/after uses the correct end position; update the code around existing,
MEMORY_SECTION_START, MEMORY_SECTION_END and the calculation of
before/after/base to use this relative find and handle the case where the end
marker is not found.
Avoids matching an unrelated end marker that appears before the start marker in CLAUDE.md. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
1. validate_no_symlink: check parent directory when destination path does not yet exist, preventing symlink-based directory traversal on first launch before fs::write creates the file. 2. inject_agent_memory: search for MEMORY_SECTION_END relative to the found start position (existing[start..]) instead of skipping past the start marker, ensuring correct absolute offset computation. 3. run_all: propagate inject_agent_memory errors with ? instead of silently dropping them via unwrap_or_else and a warning log. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Contributor
Author
|
@coderabbitai review |
✅ Actions performedReview triggered.
|
nutt-adam
added a commit
that referenced
this pull request
Mar 19, 2026
…choir runs (#71) * feat(health): add startup grace window to wait_for_agent_idle (#67) Prevent fresh prompt steps from falsely completing before the agent has consumed the prompt. The startup grace period (default 30s) gates completion detection until real working activity is observed. Key changes: - wait_for_agent_idle accepts a startup_grace Duration parameter - AgentStatus::Working counts as activity even without pane hash change, requiring 2+ consecutive polls to avoid flicker false positives - First pane capture no longer counts as a hash "change" - Completion signals before any activity are held until grace expires - "Unravelling" added to claude-code working patterns - startup_grace_secs field threaded through config and automation Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: add 0.3.0 changelog entry for issue #67 and prior unreleased changes Cover startup grace window (#67), persistent memory (#62/#63), merge gate enforcement (#59), permissions suggest (#53), orchestration state machine (#54/#55), and all fixes shipped since 0.2.0. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: reduce startup grace to 10s and validate wait settings - Reduce DEFAULT_STARTUP_GRACE_SECS from 30 to 10 so the completion-before-activity path fires before typical wait timeouts - Validate that wait_timeout_secs/startup_grace_secs are only set when wait_for_idle is true, failing fast with actionable guidance Addresses CodeRabbit feedback on PR #71. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This was referenced Mar 20, 2026
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.
Summary
memory: Option<String>field toAgentConfigfor pointing to a persistent memory filett up, claude-code agents get memory contents appended toCLAUDE.mdin their worktree under# Agent Memoryensure_tutti_dirnow creates.tutti/memory/directory..traversalTest plan
AgentConfigparses with memory field presentAgentConfigparses without memory field (backward compat, defaults to None)..)ensure_tutti_dircreates.tutti/memory/directoryprepend_memory_to_promptadds context for non-Claude runtimesprepend_memory_to_promptis a no-op for claude-code (handled via CLAUDE.md)Closes #62
🤖 Generated with Claude Code
Summary by CodeRabbit
New Features
Improvements
Tests
Documentation