Harvey Pipeline — Implementation Plan

See pipeline-design.md for the full design rationale.

New Module Dependencies

None. The pipeline is implemented entirely with existing Harvey packages and the Go standard library.

Files to Create

Files to Modify

Implementation Phases

Phase 1 — Argument Parsing

func parsePipelineArgs(root string, args []string) (threshold float64, files []string, err error)

• Validate args[0] matches \d+(\.\d+)?%; parse to float64 in (0, 1]
• Remaining tokens are workspace-relative file paths
• Each path is validated with resolveWorkspacePath (path escape blocked)
• Return error on missing files or zero paths

Phase 2 — File Reading and @mention Extraction

func readPipelineFile(root, relPath string) (body string, err error)
func scanAtMention(body string) string  // returns "" if none found

• readPipelineFile resolves and reads; enforces workspace boundary
• scanAtMention applies regex @([\w:.-]+) and returns the first capture group

Phase 3 — Model Resolution

func resolvePipelineClient(a *Agent, mention string) (LLMClient, error)

• Empty mention → return a.Client, nil (no allocation)
• Non-empty → construct temporary AnyLLMClient using same provider as a.Client but with model name = mention
• If construction fails or model unavailable → return nil, descriptive error

Phase 4 — Confidence Extraction

func extractConfidence(ctx context.Context, client LLMClient, response string) (score float64, stripped string, method string, err error)

1. Try {"confidence": X.X ...} JSON at end of response; strip block
2. If fail: send follow-up "Rate your confidence 0.0–1.0. Reply only: CONFIDENCE: <score>"; parse reply
3. If fail: keyword scan (hedging phrases → 0.30; no hedging → 0.80)

Returns the score, the response with the confidence block removed, the method used (“json”, “followup”, “keyword”), and any error from the follow-up call.

Phase 5 — Step Execution

func runPipelineStep(
    ctx context.Context,
    a *Agent,
    client LLMClient,
    messages []Message,
    out io.Writer,
    stepNum, total int,
    filename string,
    threshold float64,
) (strippedResponse string, confidence float64, err error)

1. Append hidden confidence instruction to the last User message
2. Update spinner: [N/total] filename | context X%
3. client.Chat(ctx, messages, out) — streams to terminal
4. extractConfidence on result
5. Print per-step result line with ✓ or ✗
6. If score < threshold → return error
7. Return stripped response for next step

Phase 6 — Pipeline Orchestrator

func cmdPipeline(a *Agent, args []string, out io.Writer) error

1. parsePipelineArgs — validate threshold and files
2. For each file: readPipelineFile → scanAtMention → resolvePipelineClient
   • Stop on any resolution error before running any steps
3. Fountain: open pipeline scene block
4. Step 1: messages = a.History + user message from file
5. Step N (N > 1): messages = system prompt + user message (prev response + file body)
6. runPipelineStep for each step; stop on error
7. On success: a.AddMessage("assistant", finalStrippedResponse)
8. Fountain: CUT BACK TO: original scene

Phase 7 — Tab Completion

In buildCompleter switch statement, add:

case "/pipeline":
    pathStart = 2  // tokens[1] is confidence %, paths start at index 2

Phase 8 — Tests (pipeline_test.go)

Acceptance Criteria

• /pipeline 90% step1.md step2.md runs end-to-end with mocked LLM client
• Confidence below threshold stops at the correct step
• Unresolved @mention stops immediately before any steps execute
• a.Client and a.History unchanged on any failure
• Final response appended to History on success
• Tab completion offers file paths for argument positions 2 and beyond
• Spinner shows filename and context percentage during each step
• Per-step result line printed after each step completes
• go test ./... passes