implement token reduction with headroom#12
Conversation
![sergereview[bot]](https://avatars.githubusercontent.com/in/3673521?s=60&v=4){kind=small}
There was a problem hiding this comment.
🤗 Serge says:
The PR correctly wires headroom context compression into every LLM call path, so all three modes are functionally covered (GitHub Action → action_runner.py, GitHub App webhook → app.py, staged web app → webapp.py). The compressor is attached to the single ChatCompletionClient instantiation point in reviewer.py, which means run_review, run_followup, prepare_review, and the synthesis helper all compress before posting.
However, the Action interface is incomplete: action.yml only exposes 3 of the 7 documented headroom settings as inputs, and the run: block only forwards those 3. Action users who read the README table cannot actually set HEADROOM_COMPRESS_SYSTEM_MESSAGES, HEADROOM_PROTECT_RECENT, HEADROOM_MIN_TOKENS, HEADROOM_KOMPRESS_MODEL, or HEADROOM_MODEL_LIMIT via with:. They would have to fall back to workflow-level env: variables, which contradicts the README’s “Env var / input” column.
Action interface gaps
action.ymlneeds inputs (andrun:forwarding) for the missing headroom settings so Action users can configure them the same way App/Web operators can via env vars.
Documentation gaps
.env.exampledoes not mention anyHEADROOM_*variables, making the feature undiscoverable for App/Web deployments.
Tests
tests/test_compression.pyhas good coverage of the fallback paths (missing package, disabled, empty messages, failure, model forwarding). No issues there.
Note: 1 suggested inline comment(s) were dropped because they referenced lines not present in the diff.
26 LLM turns · 36 tool calls · 1007.9s · 560483 in / 58026 out tokens
| headroom_compress_user_messages: | ||
| description: 'Set to "true" to also compress user messages (the annotated diff). Off by default so cited line numbers stay intact.' | ||
| required: false | ||
| default: 'false' |
There was a problem hiding this comment.
The run: block only forwards three headroom env vars, but the README documents seven. Action users cannot configure HEADROOM_COMPRESS_SYSTEM_MESSAGES, HEADROOM_PROTECT_RECENT, HEADROOM_MIN_TOKENS, HEADROOM_KOMPRESS_MODEL, or HEADROOM_MODEL_LIMIT via with: inputs because there are no corresponding action inputs.
Please add the missing inputs and forward them in the run: step so the Action interface matches the README table.
| HEADROOM_COMPRESS: ${{ inputs.headroom_compress }} | ||
| HEADROOM_TARGET_RATIO: ${{ inputs.headroom_target_ratio }} | ||
| HEADROOM_COMPRESS_USER_MESSAGES: ${{ inputs.headroom_compress_user_messages }} | ||
| run: reviewbot-action |
There was a problem hiding this comment.
Missing forwarding for HEADROOM_COMPRESS_SYSTEM_MESSAGES, HEADROOM_PROTECT_RECENT, HEADROOM_MIN_TOKENS, HEADROOM_KOMPRESS_MODEL, and HEADROOM_MODEL_LIMIT.
There was a problem hiding this comment.
🤗 Serge says:
This PR adds optional context compression via the headroom-ai package. The implementation is clean, well-tested, and follows safe fallback patterns: compression is opt-in, import failures and runtime errors are caught and silently fall back to uncompressed messages, and the Action only installs the heavy extra when the input is enabled. Documentation, env examples, and Action inputs are all consistent. No correctness or security issues found.
1 LLM turn · 0 tool calls · 607.2s · 11497 in / 8545 out tokens
1 participant
Fixes #6
I am guessing we need to propagate the changes to the UI as well as that is the only way to use the bot at the moment?