repo_reconstruct / swe_flow¶
SWE-Flow creates reconstruction tasks from a working repository's execution dependencies. It does not need PR history.
The September expansion completed 100 Harbor tasks across seven Python repositories (24 retained, 76 new). All 76 new exports have hash-matched baseline reward 0 and reference reward 1. Expansion cost was \(22.91**, approximately **\)0.30 per new export, including failed attempts and estimated worker/build costs. Two pilot instructions retain known specification issues: published copies carry evaluation.status = "needs_repair", with the original exports preserved. The published manifest records source diversity, costs and both findings. These generation checks do not establish independent quality acceptance for the collection. The complete bundles are published as HuggingEnvs/repo2rlenv-swe-flow.
Pipeline, step by step¶
flowchart TD
S["Pinned healthy Python repo"] --> T["Bounded runtime traces of sampled tests"]
T --> G["Group tests by observed function dependencies"]
G --> D["Order development steps; choose undeveloped functions"]
D --> C["Stub entries / remove helpers; measure test contrast"]
C --> P1["P1 · Behavioral docstrings for scheduled functions"]
P1 --> P2["P2 · Requirements from test evidence"]
P2 --> B["Skeleton repo + original-source reference + private tests"]
B --> R["Fresh Harbor nop + oracle"]
R -->|"0 / 1"| H["Export reconstruction task"]P1, P2, … identify actual model calls. Unlabelled stages are code or remote execution.
Observe dependencies. Runtime traces collect which supported functions each sampled test uses. Incomplete traces are excluded; groups are ordered by dependency count.
Choose a development step. A task introduces only not-yet-developed functions. Entry points keep signatures; new helpers are removed. The full configured suite checks healthy and skeleton states.
Write two views of the requirement. P1 consumes scheduled function source. P2 receives those generated public docstrings, the complete scheduled functions and test evidence, including parametrization decorators. It reconciles public behavior and limits requirements to the missing functions. The generated docstrings are checked for exactly one entry per scheduled node_id before insertion.
Every prompt and its data¶
Two author calls per candidate: docstrings and specification. Tracing and dependency scheduling use no LLM.
| Call | System prompt composition | User / input material | Output | Retry or branch |
|---|---|---|---|---|
| P1 · Docstrings | docstring_prompt.md + first two docstring demonstrations + adaptation | candidate.functions, indexed by node_id. | Docstrings: functions[{node_id, docstring}] | Names must match scheduled nodes exactly once. |
| P2 · Specification | specification_prompt.md + first two specification demonstrations + adaptation | candidate.test_evidence, P1 public docstrings, complete scheduled functions and scope constraints. | Specification: markdown | A separate call; bounded correction rejects private fixture/test references. |
Read the complete swe_flow prompt reference for every retained template, appended instruction, substitution, example and output schema. The shared prompt guide explains how to inspect the fully resolved request from a real run.
Follow one task¶
Illustration: tests exercise a public iterator helper and two internal helpers. A development step removes the newly required implementations, retains the rest of the repository, and asks the learner to implement the specified behavior.
What repeats, what is checked¶
Trace limits are trace_max_tests and trace_seed; no model repairs a failed trace. Authoring or Harbor failures skip the candidate. This profile does not loop back and rewrite the schedule after a failed solver rollout.
An exported bundle is a generation result. Independent leakage review, shortcut probes and blind solver traces belong to the later quality campaign.
Implementation map¶
Run and supported profile¶
Run repo2rlenv generate --config examples/owned-swe-flow.yaml. Cloud workers, campaign budgets, resume receipts and the Rich/JSON CLI use the shared owned recipe interface.
The first profile supports top-level synchronous Python functions in supplied source paths. Tracing samples 128 tests with seed 24 by default. Each traced test has a five-second profiling window; incomplete traces do not enter scheduling. The healthy and skeleton states still run the complete configured test suite. Additional threads, subprocesses, classes and asynchronous functions require another tracing profile.
Tests with the same observed function dependencies form a group. Groups are ordered by dependency count; each step introduces only previously undeveloped functions. Entry points retain their signatures and receive generated behavioral docstrings. Newly introduced helper implementations are removed. The reference restores the original source. The remaining repository supplies realistic context.
Two separate model stages generate function docstrings and test-grounded task requirements using the upstream prompts. Fresh Harbor checks establish an unsolved baseline and working reference. The initial 20-task target has expanded to 100 generated tasks; independent quality review, attack checks and model rollouts come later.
Options extend the Python build/source/test profile with target, max_candidates, trace_max_tests and trace_seed. Every trace, schedule, source snapshot, model request and trial stays in the campaign evidence directory.
Credit: SWE-Flow, MIT, commit 7da5b046fa1dc184674e4e94a9989be56c39e4e7, and SWE-Flow-Trace. See RFC 0016 and the packaged recipes/swe_flow/provenance.md for implementation differences.