push and pull_request both fired for commits on a branch with an open
PR. Scope push to master only, matching sporah's workflow, so branch
commits trigger just the pull_request run.
The docker-labeled runner's image is node:20-bookworm, needed for the
checkout/cache actions (both Node-based). Overriding it with
container: rust:1-bookworm dropped Node from the image entirely, so
checkout failed with "node: executable file not found in $PATH". The
rust-ci label (see sporah's workflow) points at a custom image with
both Rust and Node preinstalled, avoiding the conflict.
deep_research is the only project this repo is meant to showcase, so the
Cargo workspace wrapping it and an unrelated side project no longer earns
its keep:
- swear_cleanup moved to a new standalone local repo (~/dev/swear_cleanup,
not pushed anywhere) via `git subtree split`, with its pre-workspace-
split history (when it lived at src/swear_cleanup/ in a single shared
crate) spliced onto its post-split history rather than starting from a
single flattened snapshot. FINDINGS.md, which was sitting at this repo's
root but was actually swear_cleanup's own build log, went with it.
- deep_research/{src,Cargo.toml,README.md,docs} moved to the repo root;
the [workspace] table collapsed into a plain [package] manifest with
dependency versions inlined from the old [workspace.dependencies].
- Cargo.toml keeps an explicit empty [workspace] table (not just omitted)
so that checking this repo out as a nested git worktree — this
project's own normal workflow — can't accidentally inherit a stale
ancestor directory's workspace manifest, which is exactly what broke
the build while testing this change from a worktree.
- .forgejo/workflows/deep_research-ci.yml -> ci.yml, dropping the now-
meaningless -p deep_research scoping and path filters (redundant when
it's the only thing in the repo).
- README.md and docs/case-study.md updated for the flattened commands
(cargo run/test with no -p flag); their relative links to each other
and to src/ were already correct since both moved together.
Verified: cargo build/test/clippy/fmt all clean from the new repo root.
- .forgejo/workflows/deep_research-ci.yml: build, test, clippy (-D
warnings), and fmt --check on push/PR, scoped to deep_research (not
workspace-wide — swear_cleanup has an unrelated pre-existing clippy
warning that would otherwise break CI on an unrelated project)
- README.md: what the project does, the four-agent architecture, why
it's local-first (Ollama + self-hosted SearXNG, no cloud API key, no
query leaves the host), project layout, and how to run/test it
- docs/case-study.md: narrative walkthrough of the max-turns recovery
path, the DuckDuckGo-rate-limiting root cause and SearXNG fix, and the
separation-of-concerns refactor — each step verified against a live
run of the actual failing case, not just unit tests. Uses a neutral
"AI customer-support chatbot trends" research run as the illustrative
clean-pipeline example rather than the personal topic used during
actual debugging.
core.rs had grown into a 612-line grab-bag mixing six unrelated concerns:
CLI arg parsing, logging setup, top-level orchestration, the researcher
agent phase, chat-history reconstruction utilities, the summarizer agent
phase, and the writer agent phase — while review.rs, tools.rs, stream.rs,
and progress.rs already correctly isolated their own concerns. This
splits core.rs to match that existing pattern instead of being the one
file that doesn't follow it:
- cli.rs — Cli struct + DEFAULT_TOPIC
- observability.rs — initialize_observability
- models.rs — RESEARCHER_MODEL / WRITER_MODEL (previously duplicated
across call sites, now a single source of truth)
- history.rs — pure chat-history parsing/reconstruction helpers
(partial_findings_from_history, annotated_transcript_from_history, and
their private helpers), plus their unit tests. Also dedupes
MAX_TOOL_RESULT_CHARS, which was previously defined twice.
- researcher.rs — gather_findings + GatheredFindings (the tool-calling
research phase)
- summarizer.rs — summarize_partial_history (the max-turns recovery
agent)
- writer.rs — write_report
- research.rs — the top-level research() orchestration loop
main.rs now only does argument parsing, logging setup, and the top-level
call — no orchestration logic of its own. Unit tests stay co-located
with the code they test per Rust convention (not pulled into separate
files) rather than under "prefer new files" — that applies to
production code organization here.
No behavior changes; cargo test/clippy/fmt all clean.
DuckDuckGo's HTML endpoint rate-limits after enough requests, and a
rate-limited response is indistinguishable from a genuine empty result —
which is exactly what burned a full 12-turn research run on 13 consecutive
"No results found" responses. Swapping to a local SearXNG instance's JSON
API (no HTML scraping needed) fixes both problems: SearXNG spreads queries
across multiple upstream engines instead of hammering one, and this
machine already runs an instance.
This tool is explicitly local-only and never released, so the base URL is
a plain default (localhost:8080) overridable via SEARXNG_URL, not a
general-purpose config surface. Evaluated the two third-party SearXNG
crates on crates.io first (searxng, searxng-client) — both are
single-maintainer v0.1.0 packages with no adoption signal and no official
alternative exists, so a hand-rolled reqwest + serde call was the better
bet for something this small.
Drops the DuckDuckGo-specific HTML parsing (parse_search_results,
resolve_ddg_redirect, the .result/.result__a/.result__snippet scraper
selectors) entirely — fetch_page's extract_readable_text still needs
scraper for arbitrary fetched pages, so that dependency stays.
Adds an #[ignore]'d live smoke test (search_web_returns_real_results_from_local_searxng)
for manually verifying against a running instance; not run by default
since there's no CI environment with SearXNG available.
Add summarize_partial_history: a one-shot writer-model pass that turns an
annotated transcript (tool calls with their args, so a fetch's URL or a
search's query stays attached to its result, plus results and interim
notes) into a proper footnote-style findings dump, instead of the flat
concatenation partial_findings_from_history produces on its own.
It's wired in as the MaxTurnsError recovery path in gather_findings, but
partial_findings_from_history stays as the fallback for an empty transcript
or if the summarizer call itself fails — the one guaranteed recovery path
shouldn't have a second turn-budget/model failure as a single point of
failure.
Observability: instrumented with the same #[tracing::instrument(fields(
gen_ai.agent.name = ...))] + spinner pattern as the researcher/reviewer/
writer phases, with info!/warn! events on success, empty-transcript
skip, and summarizer failure.
Also: cargo fmt across the crate (unrelated formatting drift had
accumulated), and adds unit tests for the new transcript_lines /
annotated_transcript_from_history helpers.
Covers truncate, assistant_text, tool_result_text, and
partial_findings_from_history in isolation (fallback text, ordering,
truncation, and filtering out non-text content). gather_findings/
write_report/research still need a live ollama client and aren't covered
here.
Split into two filter_map-based helpers (assistant_text, tool_result_text)
and a shared truncate() so the extraction reads as a flat iterator chain
instead of four levels of nesting.
When the researcher agent exhausts its turn budget mid-investigation, rig
now surfaces PromptError::MaxTurnsError with the chat history intact rather
than nothing at all. Catch it, reconstruct a findings dump from whatever
assistant text and tool results the run produced, and still write a report
from that — clearly flagged as incomplete — instead of propagating the
error and losing all the work.
Between sending the report prompt and the writer actually starting to
generate, the spinner was already dropped, so the terminal went blank
for however long that gap was. Thread the spinner into
write_text_stream instead and stop it right as the first chunk
arrives, covering the wait with the same "Writing report..." line
rather than clearing it early.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Tag each phase (research, web search, page fetch, review, rejection,
report writing) with a distinct emoji so the spinner line shows what's
happening at a glance. Search/fetch tool calls now update the active
spinner's message directly via a small shared handle, since they run
as plain tool functions without one threaded down to them otherwise.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
write_text_stream() in the new stream.rs doesn't touch anything specific
to write_report (topic, findings, the agent) — it just drains a
MultiTurnStreamItem stream, writes each text chunk to a caller-provided
writer, and returns the accumulated string. Pulling it out lets it be
covered by unit tests against a mocked stream and an in-memory writer,
independent of a live model.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
No destructuring needed since only .text is used; matching the whole
Text struct and drops the now-unused rig::message::Text import.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
print!/println! each acquire stdout's lock internally; doing that per
streamed chunk in a tight loop adds needless contention. Lock once up
front and write!/writeln! through the held handle instead — which also
means the trailing newline must go through that same handle rather than
println!, since re-locking from the same thread would deadlock.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
print! already locks stdout per call, so the manual lock()/write!() was
extra ceremony over what the flush actually needed. Matches rig's own
cli_chatbot streaming example more closely.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Wires up starter.rs -> core.rs (CLI parsing and observability init moved
into core, main.rs left as a thin entry point) and switches the report
phase from Agent::prompt to rig's stream_prompt, printing each text
delta to stdout as it arrives instead of waiting for the full response.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
deep_research and swear_cleanup were sharing one Cargo.toml, so every
build compiled clap/indicatif/scraper/chrono (only needed by
deep_research) even when just building swear_cleanup for its own
course work, and vice versa. Moves each into its own workspace member
crate (deep_research/, swear_cleanup/) with an independent Cargo.toml
declaring only the deps it actually uses; common deps/versions are
pinned once via [workspace.dependencies] so the two don't drift.
Verified `cargo build -p swear_cleanup` alone no longer pulls in
clap/indicatif/scraper/chrono (schemars still compiles for it, but
that's a direct transitive dependency of rig itself, not something
this split can avoid). Also verified the relocated deep_research
binary still runs end-to-end against live Ollama with correct
footnote citations and sources.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Adds a clap-derive Cli (topic argument, --log-level flag) so the
research loop is a proper command-line tool instead of a raw
env::args().nth(1) read. Logging is now opt-in: tracing only
initializes a subscriber when --log-level is passed, so the terminal
stays clean by default. When logging is off, each research phase
(researcher/reviewer/writer) shows an indicatif spinner instead, so
the user isn't staring at a blank terminal during the 1-3 minute
Gemma tool-calling turns.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Interpolates today's date into the researcher's preamble so it can judge
source freshness instead of relying on training-cutoff knowledge, and
asks it to cite facts with bracketed footnote numbers backed by a
Sources list, which the writer agent is now instructed to preserve
through to the final report.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Adds a reviewer step (gemma4-e4b, fresh context) between gathering and
writing: it uses rig's typed Extractor to judge whether the findings'
conclusions actually follow from their cited sources, rather than
relying on free-text parsing. research() is now a plain bounded loop —
"the least agentic design that solves the problem", per rig's own
workflow guidance — that reruns the researcher with the reviewer's
solid_findings/gaps feedback folded into the next round's task until
it approves or MAX_RESEARCH_ROUNDS runs out.
Agents were showing up as "Unnamed Agent" in rig's built-in gen_ai.*
spans. Naming them via .name(...) fixes that, and splitting the two
phases into #[tracing::instrument]-annotated functions wraps rig's
per-turn chat/execute_tool spans in a parent span per phase, making
the trace tree legible instead of a flat stream of chat calls.
Pins rig to the newest published crates.io release (0.41.0) instead of
the git main branch, and adapts swear_cleanup's revise.rs to that
release's API (OneOrMany::first() returns T directly, raw_completion
folded into CompletionResponse::raw_response).
The research agent (gemma4:26b) drives rig's AgentRunner tool-calling
loop with two lean #[rig::tool_macro] tools — a DuckDuckGo HTML search
and a page-text fetcher — to gather and cross-check findings. A second
agent (gemma4-e4b) turns those findings into a structured report; the
smaller/faster model suffices there since it's reformatting already-
digested notes rather than doing multi-step research reasoning.
Gemma now checks its own output (a fresh, stateless completion call,
not conversation history) before it's accepted as a seed or revision,
retrying up to 5 times if it's a refusal, meta-commentary describing
what it's about to write, or a list of multiple options instead of
one direct answer. Originally tried a separate CPU-only judge model
(critic-cpu) to avoid VRAM contention, but it was both far slower
(13-16s per judgment vs Gemma's own sub-second calls) and unreliable
on the exact failure patterns it was meant to catch — Gemma reviewing
itself turned out faster and more accurate, so critic-cpu is dropped
entirely.
Also required two rounds of prompt tuning, driven by live failures:
first adding concrete negative examples after the judge approved
outputs it should have rejected, then explicitly scoping the check to
format only after Gemma started rejecting its own genuinely hostile
(but well-formed) output — conflating "should I have generated this"
with the format question actually asked. Added integration tests
covering both directions (rejecting bad formats, accepting hostile-
but-well-formed content) as a fast regression check against an
expensive full generation loop.
Enable Rig's built-in tracing spans (model, token usage, cache hits,
latency) via tracing-subscriber, filterable through RUST_LOG and
defaulting to info level. Logs write to stderr so stdout stays
reserved for program output. Standardizes the remaining ad-hoc
println! diagnostics (server startup, per-iteration revision progress,
non-convergence) into structured tracing events at appropriate levels.
Gemma now seeds deliberately hostile text, Shieldstral scores it, and
Gemma revises its own output based on the score until it drops below a
safety threshold (or a max-iteration cap is hit, returning the best
attempt seen). Extracted into a new revise module: score() now borrows
instead of consuming its args so it can run repeatedly, and the
gemma-call/extract-text logic is shared between seed generation and
every revision instead of being duplicated.
Also adds FINDINGS.md logging what actually turned out to be real
obstacles vs. overblown vs. irrelevant while building this out.
Checks whether llama-server is already healthy on startup and spawns it
from configured binary/model paths if not, polling until ready. Server
infra config (binary, model path, host, port, context size) split out
of prompts.toml into its own server.toml, and all of it lives in a new
server module rather than inline in main.rs, alongside a single reused
HTTP client and a shared health-check helper. Gemma client setup now
runs concurrently with the server health-check/spawn since they're
independent.
