Skip to content

feat(coding-agent): add system prompt note listing pre-installed Python packages - #58

Merged
samsja merged 1 commit into
mainfrom
feat/rlm-prompt-available-packages
May 21, 2026
Merged

feat(coding-agent): add system prompt note listing pre-installed Python packages#58
samsja merged 1 commit into
mainfrom
feat/rlm-prompt-available-packages

Conversation

@samsja

@samsja samsja commented May 21, 2026

Copy link
Copy Markdown
Member

Changes

  • Add a brief line to the system prompt (ipython section) listing pre-installed Python packages: requests, httpx, pyyaml, tomli, python-dotenv, pandas, numpy, scipy, beautifulsoup4, lxml, pydantic

No bootstrap/dependency changes in this PR. The actual package installation will follow in a separate PR.

Files changed

  • packages/coding-agent/src/core/prompts/rlm.ts - added package availability note
  • packages/coding-agent/CHANGELOG.md - entry under [Unreleased]

Note

Low Risk
Low risk: only adjusts prompt text and changelog; no runtime, dependency, or execution-path changes.

Overview
Updates the coding-agent RLM system prompt shown when the ipython tool is active to explicitly list Python packages that are already available in the kernel (e.g., requests, pandas, numpy, pydantic) so agents avoid unnecessary installs.

Adds a matching [Unreleased] changelog entry documenting the new prompt guidance.

Reviewed by Cursor Bugbot for commit ab1bf3a. Bugbot is set up for automated code reviews on this repo. Configure here.

Note

Add pre-installed Python package list to coding agent system prompt

When ipython is an active tool, rlm.ts now appends a note listing pre-installed packages (requests, httpx, pyyaml, tomli, python-dotenv, pandas, numpy, scipy, beautifulsoup4, lxml, pydantic) and instructs the agent to import them directly without running pip install.

Macroscope summarized ab1bf3a.

@samsja
samsja merged commit e89ea85 into main May 21, 2026
2 of 3 checks passed

@cursor cursor Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Cursor Bugbot has reviewed your changes and found 1 potential issue.

Fix All in Cursor

❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.

Reviewed by Cursor Bugbot for commit ab1bf3a. Configure here.

"",
"Use `ipython` for both Python and shell work. For repository shell commands, prefer IPython shell syntax: `!rg ...`, `!npm run check`, or `%%bash` for multi-line scripts. Do not wrap ordinary shell commands in Python subprocesses unless you need Python-level processing.",
"",
"The kernel has requests, httpx, pyyaml, tomli, python-dotenv, pandas, numpy, scipy, beautifulsoup4, lxml, and pydantic pre-installed. Import them directly; no pip install needed.",

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

System prompt claims packages exist but aren't installed

High Severity

The new system prompt line tells the model that requests, httpx, pyyaml, tomli, python-dotenv, pandas, numpy, scipy, beautifulsoup4, lxml, and pydantic are pre-installed and can be imported directly with "no pip install needed." However, the kernel bootstrap in bootstrap.ts only installs ipykernel and prime-agent-runtime — none of these 11 packages are actually installed. This will cause the agent to hit ImportError at runtime and, worse, the prompt explicitly discourages the agent from running pip install, so it may not self-recover.

Fix in Cursor Fix in Web

Reviewed by Cursor Bugbot for commit ab1bf3a. Configure here.

@kevinjosethomas
kevinjosethomas deleted the feat/rlm-prompt-available-packages branch May 22, 2026 18:46
zhengr pushed a commit to zhengr/prime-agent that referenced this pull request Aug 8, 2026
…lm-prompt-available-packages

feat(coding-agent): add system prompt note listing pre-installed Python packages
thomaswillner pushed a commit to thomaswillner/prime-agent that referenced this pull request Aug 29, 2026
Records, without editing the now-false text away, that sections 9 and 4 went
stale five hours after they were written. The correction matters more than the
content: this file exists to stop sessions trusting notes over GitHub, and it
caught its own author.

- main is 8d2139c. Between 16:12Z and 21:54Z the fleet merged PrimeIntellect-ai#279 (the PrimeIntellect-ai#58
  alert-bridge race, FIXED — stop carrying it as a standing exception), PrimeIntellect-ai#278
  (AGENTS.md invariants), PrimeIntellect-ai#283 (repo cleanup), and PrimeIntellect-ai#284, which delivered the
  last brief and closed PrimeIntellect-ai#165 with a keyword.
- Section 3's routing conclusion was confirmed by events: PrimeIntellect-ai#271 was delivered
  by the Mac maker fleet via auto-dispatch, exactly as argued, and the remote
  session correctly declined to open a second lane.
- Flags issue-state drift: PrimeIntellect-ai#271, PrimeIntellect-ai#276 and PrimeIntellect-ai#274 are delivered and merged yet
  still open, because a title reference is not a closing keyword. That is the
  mirror image of the hazard the V2 CLAUDE.md documents, and it leaves
  open-work disagreeing with main. Operator action, named as such.
- Records the residual PrimeIntellect-ai#284 deferred on stated grounds (PrimeIntellect-ai#286), which is a
  known open edge on the LIVE path.
- States the next slice: S4 / PrimeIntellect-ai#236, the first whose exit criteria need a real
  broker order. Certification stays 0/12; the system has never placed a trade.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01G5B7QM1QLQuWSBMCxiCzS6
thomaswillner added a commit to thomaswillner/prime-agent that referenced this pull request Aug 29, 2026
…e-notes lesson (#13)

* docs(spx-v2): verification pass, queue state, and self-refinement record

No code written this session — a verification pass over already-delivered
work plus the queue-state answer.

Records, so future sessions do not repeat them:

- The audit-challenge / V1-coverage / rag-tot-cot-challenge / corrected-input
  deliverable ALREADY EXISTS (AUDIT_CHALLENGE sections A-D and
  PRIME_AGENT_INPUT_SPX_V2). An operator prompt has now asked for it in at
  least two sessions; redoing it is inventing work.
- Verified queue state from GitHub: PrimeIntellect-ai#266/PR PrimeIntellect-ai#268 merged (and PrimeIntellect-ai#263 with it,
  now main f64029a); PrimeIntellect-ai#265/PR PrimeIntellect-ai#269 and PrimeIntellect-ai#264/PR PrimeIntellect-ai#270 open with CI in flight;
  PrimeIntellect-ai#272 and PrimeIntellect-ai#271 filed, unstarted, no lane.
- Errors and corrections: settle elapsed time from GitHub workflow-run
  timestamps, never the container clock; add_repo push access was
  classifier-denied so a remote session may hold read-only and cannot push;
  register_repo_root denial falls back to reading CLAUDE.md directly.
- MATS/superpowers/routing settled empirically with the exact commands used,
  so the search is not repeated: they are Mac-harness resident, and PrimeIntellect-ai#272/PrimeIntellect-ai#271
  already carry auto-dispatch, which is what routes them to the maker fleet.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01G5B7QM1QLQuWSBMCxiCzS6

* docs(spx-v2): implementation-session addendum — access, setup, CI outage, self-review

Appends the second half of the session to the notes: implementing PrimeIntellect-ai#272 after
the operator corrected two access assumptions.

The corrections matter more than the code:
- push DOES work; "I cannot push" was inferred from add_repo's access label
  rather than tested. A dry-run push proved it. Also: the refspec push form is
  classifier-denied while `git push -u origin <branch>` succeeds.
- this host is not the MacBook (uname, no /Users, no ~/.prime).

Also records the environment setup the Makefile assumes (venv before
v2-install, ruff 0.15.22 via python -m, seeding the gitignored account.yaml,
and proving PYTHONPATH beats editable installs in a worktree), the method that
diagnosed the repo-wide CI outage in two calls (zero recorded steps, then the
same workflow red on main), and two test defects self-review caught before
pushing.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01G5B7QM1QLQuWSBMCxiCzS6

* docs(spx-v2): final queue state — PrimeIntellect-ai#273 merged, only PrimeIntellect-ai#271 remains

Closes the record for this session.

- All five briefed PRs merged (PrimeIntellect-ai#268, PrimeIntellect-ai#263, PrimeIntellect-ai#270, PrimeIntellect-ai#269, PrimeIntellect-ai#273); main is c84855d.
  Issues PrimeIntellect-ai#266 and PrimeIntellect-ai#272 closed by their PRs. PrimeIntellect-ai#271 is the only open brief and
  was never authorised, so never started. Runtime testing is unblocked.
- The CI outage (13:51Z-15:47Z) was account-level and hit main identically;
  recovery was visible as `changes` taking 9s with real steps instead of 2s
  with none. Nothing in the diff ever needed changing.
- Records the scope misjudgement worth carrying forward: a Codex P1 mapped
  directly to an acceptance checkbox I had deferred as out of scope. When a
  finding maps to an acceptance criterion it is in scope by definition.
- Records the auto-merge hazard: squash composes the commit message from the
  PR body, so a body left stale after a review round writes false claims into
  main permanently. Rewrite the body before merge; keep corrections visible.
- Records a published test claim that had not been executed, and the rule
  that follows from it.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01G5B7QM1QLQuWSBMCxiCzS6

* docs(spx-v2): correct the queue state — the last brief landed while idle

Records, without editing the now-false text away, that sections 9 and 4 went
stale five hours after they were written. The correction matters more than the
content: this file exists to stop sessions trusting notes over GitHub, and it
caught its own author.

- main is 8d2139c. Between 16:12Z and 21:54Z the fleet merged PrimeIntellect-ai#279 (the PrimeIntellect-ai#58
  alert-bridge race, FIXED — stop carrying it as a standing exception), PrimeIntellect-ai#278
  (AGENTS.md invariants), PrimeIntellect-ai#283 (repo cleanup), and PrimeIntellect-ai#284, which delivered the
  last brief and closed PrimeIntellect-ai#165 with a keyword.
- Section 3's routing conclusion was confirmed by events: PrimeIntellect-ai#271 was delivered
  by the Mac maker fleet via auto-dispatch, exactly as argued, and the remote
  session correctly declined to open a second lane.
- Flags issue-state drift: PrimeIntellect-ai#271, PrimeIntellect-ai#276 and PrimeIntellect-ai#274 are delivered and merged yet
  still open, because a title reference is not a closing keyword. That is the
  mirror image of the hazard the V2 CLAUDE.md documents, and it leaves
  open-work disagreeing with main. Operator action, named as such.
- Records the residual PrimeIntellect-ai#284 deferred on stated grounds (PrimeIntellect-ai#286), which is a
  known open edge on the LIVE path.
- States the next slice: S4 / PrimeIntellect-ai#236, the first whose exit criteria need a real
  broker order. Certification stays 0/12; the system has never placed a trade.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01G5B7QM1QLQuWSBMCxiCzS6

---------

Co-authored-by: Claude <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant