Problem
BugFixInputs.comment is only honored on the Phabricator follow-up path. agents/bug-fix/hackbot_agents/bug_fix/agent.py:117-126 renders prompts/follow-up.md with {comment} when revision_id is set, and otherwise renders prompts/triage-and-fix.md, which has no {comment} placeholder:
Bug to work on: {bug_id}
Task: Triage and fix the bug, and verify the fix
The rules in {rules_path} are available if the task calls for them, but the task above is your primary directive and overrides the default triage workflow.
The AgentInputs validator only enforces the other direction (revision_id ⇒ comment), and BugFixInputs ignores unknown fields, so POST /agents/bug-fix/runs with {"bug_id": 1234567, "comment": "start from the regression range in comment 4"} is accepted, stored on the run, passed to the container as COMMENT, and then silently dropped. Nothing tells the caller their instructions went nowhere.
The consequence is that there is no way to say anything to a plain bug-fix run beyond the bug id — not from the API, not from TriggerForm.tsx (which collects only bug_id, model, max_turns, effort), and not from a Slack action button.
Change
A new optional instructions on BugFixInputs (services/hackbot-api/app/schemas.py:79-88) and on the agent's AgentInputs (agents/bug-fix/hackbot_agents/bug_fix/__main__.py:8-45) — it reaches the job as INSTRUCTIONS for free through model_to_env — plus an {instructions} slot in triage-and-fix.md: empty when unset, and otherwise framed as guidance that does not override the rules, in the <...>-tag style follow-up.md already uses for untrusted comment text. The str.format safety argument at agent.py:47-55 covers it: substituted values are inserted verbatim and not re-scanned, so an untrusted instructions cannot break out of its placeholder.
Reusing comment instead would make one field mean two different things depending on whether revision_id is set, and would leave the follow-up path's revision semantics (phabricator_update_patch / phabricator_add_comment, config.py:30-36) entangled with a plain "focus here" hint. A separate field is cheaper to reason about.
Worth a <textarea> in TriggerForm.tsx alongside it, the way AutowebcompatReproInputs.bug_data and TestPlanGeneratorInputs.feature_description already have one.
Why this is filed on its own
It came up as the second half of a Slack action-button request: one button that triggers a bug-fix run on the triaged bug, and a second that triggers it with an accompanying prompt. The second one is not possible today, and it also needs a Slack modal (views.open + view_submission, with private_metadata carrying the message identity so the trigger claim still applies), which is receiver work on a different schedule.
Keeping that separate lets this input land whenever it lands. The silent drop of comment on a plain run is worth fixing on its own merits even if a second Slack button never ships.
Problem
BugFixInputs.commentis only honored on the Phabricator follow-up path.agents/bug-fix/hackbot_agents/bug_fix/agent.py:117-126rendersprompts/follow-up.mdwith{comment}whenrevision_idis set, and otherwise rendersprompts/triage-and-fix.md, which has no{comment}placeholder:The
AgentInputsvalidator only enforces the other direction (revision_id⇒comment), andBugFixInputsignores unknown fields, soPOST /agents/bug-fix/runswith{"bug_id": 1234567, "comment": "start from the regression range in comment 4"}is accepted, stored on the run, passed to the container asCOMMENT, and then silently dropped. Nothing tells the caller their instructions went nowhere.The consequence is that there is no way to say anything to a plain bug-fix run beyond the bug id — not from the API, not from
TriggerForm.tsx(which collects onlybug_id,model,max_turns,effort), and not from a Slack action button.Change
A new optional
instructionsonBugFixInputs(services/hackbot-api/app/schemas.py:79-88) and on the agent'sAgentInputs(agents/bug-fix/hackbot_agents/bug_fix/__main__.py:8-45) — it reaches the job asINSTRUCTIONSfor free throughmodel_to_env— plus an{instructions}slot intriage-and-fix.md: empty when unset, and otherwise framed as guidance that does not override the rules, in the<...>-tag stylefollow-up.mdalready uses for untrusted comment text. Thestr.formatsafety argument atagent.py:47-55covers it: substituted values are inserted verbatim and not re-scanned, so an untrustedinstructionscannot break out of its placeholder.Reusing
commentinstead would make one field mean two different things depending on whetherrevision_idis set, and would leave the follow-up path's revision semantics (phabricator_update_patch/phabricator_add_comment,config.py:30-36) entangled with a plain "focus here" hint. A separate field is cheaper to reason about.Worth a
<textarea>inTriggerForm.tsxalongside it, the wayAutowebcompatReproInputs.bug_dataandTestPlanGeneratorInputs.feature_descriptionalready have one.Why this is filed on its own
It came up as the second half of a Slack action-button request: one button that triggers a bug-fix run on the triaged bug, and a second that triggers it with an accompanying prompt. The second one is not possible today, and it also needs a Slack modal (
views.open+view_submission, withprivate_metadatacarrying the message identity so the trigger claim still applies), which is receiver work on a different schedule.Keeping that separate lets this input land whenever it lands. The silent drop of
commenton a plain run is worth fixing on its own merits even if a second Slack button never ships.