Repository navigation
Name the tools in the files agents read first - #76
Merged
Merged
Conversation
The middle row of the benchmark table is an agent with no daemon, no injection and no wrappers, reading .continuum/ cold. It scores the same as full Continuum. Recording the context is what produces the accuracy; injection makes it fast rather than correct. Neither the README nor the command surface said so. `continuum help --all` printed all 59 commands as one flat argparse list, which reads as if worktrees and teams matter as much as `note`. Group that listing instead, recording commands first, with coordination and governance named as optional. Help strings come from the parser, so they cannot drift from the commands they describe. A test asserts set equality both ways, so a new command cannot silently fall out of the grouping and a deleted one cannot linger in it. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01G8NSqd1KmNsDBsHnWR9SRJ
The instruction block every agent gets described Continuum's tools in prose. That asks an agent to know the tool names before it has seen the MCP tool list, and gives an agent whose harness never connects the server nothing to act on at all. It now carries a table: one row per thing you need, the tool that does it, and the command that does the same thing without MCP. The commands are the real signatures — `handoff` takes both --task and --next-step, and there is no `log --id`. `get_raw_log` is marked a last resort so the compact views stay the read path. Added GitHub Copilot as an install target. It reads `.github/copilot-instructions.md` and ignores AGENTS.md, so it was the one widely-installed agent the installer left out. It ships inside VS Code and the JetBrains IDEs rather than as a CLI, so detection goes through the editor rather than through PATH. The repo's own AGENTS.md, CLAUDE.md, GEMINI.md and Claude skill were stale copies of an older, vaguer version of the block. Regenerated from the source so what Continuum tells other projects is what it tells itself. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016EpRGYJJtxUFiJrT6RywR1
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: a078fb3c84
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Review on #76 caught that the new tool table reached almost nobody. Both writers treated the presence of any Continuum block as final, so a project installed under an earlier release kept that release's instructions forever and every improvement landed only in projects installed after it. That is the worse half of the bug this PR set out to fix: instructions no one receives. The block is now delimited and carries a version derived from a digest of its own text, so an edit to the instructions cannot ship without changing the version — the failure mode of a hand-bumped constant is exactly the one above. A block written before the markers existed is found by its headings and ends where a heading Continuum did not write begins, so what you put around it, above or below, survives the upgrade. Rerunning when nothing changed still reports `already`; a real change reports `updated`. Frontmatter is passed separately rather than folded into the body. Cursor and Claude only read it as the first thing in the file, so the marker cannot sit above it. Two smaller findings from the same review: - The no-MCP fallback for "where the work stands" pointed at `continuum status`, which prints health metadata and counters and never the task or the next step. Bare `continuum` is the command that prints those. - Copilot detection checked only VS Code, while the code claimed to support the JetBrains IDEs. It now looks for a JetBrains configuration directory on each platform, so a JetBrains-only machine is no longer skipped. Full suite: 777 tests, OK. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016EpRGYJJtxUFiJrT6RywR1
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
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.
Two commits, both about what an agent knows before anyone types anything.
Say what the benchmark already shows: this is a recorder (ea6d319)
Positioning carried through to Continuum's own surface.
Name the tools in the files agents read first (a078fb3)
The instruction block every agent gets described Continuum's tools in prose.
That asks an agent to know the tool names before it has seen the MCP tool
list, and gives an agent whose harness never connects the server nothing to
act on at all. It now carries a table — one row per thing you need, the tool
that does it, and the command that does the same thing without MCP. The
commands are the real signatures:
handofftakes both--taskand--next-step, and there is nolog --id.get_raw_logis marked a lastresort so the compact views stay the read path.
GitHub Copilot is a new install target. It reads
.github/copilot-instructions.mdand ignores AGENTS.md, so it was the onewidely-installed agent the installer left out. It ships inside VS Code and the
JetBrains IDEs rather than as a CLI, so detection goes through the editor
rather than through PATH.
The repo's own
AGENTS.md,CLAUDE.md,GEMINI.mdand Claude skill werestale copies of an older, vaguer version of the block. Regenerated from the
source, so what Continuum tells other projects is what it tells itself.
Tests
python -m unittest discover -s tests— 773 tests, OK. Two new cases intests/test_integrations.py: Copilot's file is written once and isidempotent, and the shared block names both the tools and the commands.
🤖 Generated with Claude Code
https://claude.ai/code/session_016EpRGYJJtxUFiJrT6RywR1