Flox Debug

Flox Debugging

Debug a Flox environment when it doesn't behave as expected. Its current coverage is catalog resolution: why the resolver picks (or skips) specific package builds, and why adding packages to an existing environment can fail with constraint errors.

Resolution failures are usually page problems, not version-conflict problems. The reflex is to read constraints_too_tight as two packages wanting incompatible versions of a shared dependency. That is almost never what it means here — constraints_too_tight is the generic group-level failure and names no cause by itself. The cause is one of three things:

• Page coverage — no single base page carries every package in the group. The most common.
• Metadata and licence exclusions — allow.unfree, allow.broken or allow.licenses ruled a package out on every page.
• Version conflicts — a genuine clash between version constraints. The least common; reach for it last.

Do the page analysis first, check the [options] exclusions second, and fall back to version-conflict reasoning only once both rule it out.

Establish Context

Work it out yourself first. Ask only for what you cannot determine, and never open with a questionnaire.

1. Find the environment. Look for .flox/env/manifest.toml in the current directory, then in any directory the user named. If there is no manifest anywhere, you are debugging a standalone package — say so and carry on.

2. Identify the package group. Read it from the manifest: packages with no explicit pkg-group are in the implicit group, which Flox calls toplevel — its error messages say constraints for group 'toplevel' are too tight. If the manifest declares more than one group, take the one containing the package the user is complaining about and say which you picked. Every package in a group must resolve to the same base page, so the group scopes the whole diagnosis.

3. Identify the packages involved. Take the installed set from the manifest. Only a package the user is adding may need asking, and only if their message did not already name it.

Produce the diagnosis from what you can read, and state the assumptions you made. A user corrects a wrong assumption faster than they answer three questions.

Parse the Manifest

Read the manifest.toml from the environment. Extract all packages in the target package group:

# Example manifest entries:
[install]
gh.pkg-path = "gh"                        # default group
python3.pkg-path = "python3"              # default group
claude-code.pkg-path = "flox/claude-code"
claude-code.pkg-group = "claude-code"     # separate group
nodejs_22.pkg-path = "nodejs_22"
nodejs_22.version = "22.14.0"
nodejs_22.pkg-group = "nodejs"            # separate group

For each package in the target group, build a PackageDescriptor:

• install_id: the TOML key (e.g., gh, python3)
• attr_path: the pkg-path value
• systems: the environment's declared systems, not the local platform. Read [options] systems from the manifest. If a package carries its own .systems, use that list for that descriptor instead. If [options] systems is absent the environment targets the DEFAULT set — which as of Flox 1.15 is three systems (aarch64-darwin, aarch64-linux, x86_64-linux); x86_64-darwin is only targeted when declared explicitly — confirm against manifest.lock, where every locked package records its system.
• version: the version value if present, null otherwise
• Skip packages with a flake attribute — those are not resolved through the catalog

Never narrow systems to your own machine. Two of the message types below — attr_path_not_found.systems_not_on_same_page and attr_path_not_found.not_found_for_all_systems — are multi-system failures by definition. A single-system reproduction resolves cleanly against the very failure you were asked to debug, and you will report "works fine" on a broken environment. This is about the fidelity of the request, not about how to read the response — see Step 2 before treating any attr_path_not_found as evidence.

Then apply the environment's [options] to every descriptor in the group. These change what resolves, so a reproduction that omits them is not a reproduction:

Send the false rows too. allow_unfree defaults to true in the API, so omitting it is not neutral — it silently grants what the manifest denied. Verified with one descriptor twice: field omitted resolves cleanly on page 1017464; allow_unfree: false returns page: null with an error-level unfree message. Skip that row and a genuinely failing install becomes a clean reproduction.

allow_broken defaults to false and allowed_licenses defaults to unset (no restriction), so those two are safe to leave out when the manifest does not set them. allow_unfree is the only inverted default.

unfree, broken and licenses are the only keys [options].allow accepts. The API also accepts allow_insecure, allow_pre_releases and allow_missing_builds on a descriptor, but no manifest key sets them — leave them at their defaults when reproducing an environment.

The unfree and broken message types below are exactly what those two boolean flags gate (insecure is the same shape, gated by allow_insecure, which no manifest key sets). allow.licenses behaves differently: a package excluded by allowed_licenses is reported at trace as resolution_logic — e.g. TRACE (hello): The license GPL-3.0-or-later is not in the allowed licenses: ['MIT']. — and not as an unacceptable_licenses message. See Step 3 for why that matters.

Then append the user's new packages as additional descriptors in the same group.

Core Concepts

Two Kinds of Page

Every resolved package sits at the intersection of two axes:

• Base page — the nixpkgs revision the package was built against. This is the page number in the API response. Higher = newer nixpkgs.
• Build page — the source repository revision. This is rev_count / rev / rev_date in the response. Higher rev_count = newer source.

The resolver picks the highest base page where ALL packages in the group can be satisfied. A newer build (high rev_count) on an old base page can drag the entire group down — or make resolution impossible if no single base page has all packages.

Why This Causes Confusion

Scenario 1: New publish not picked up. A user publishes a new version. The publish succeeds. But flox install still picks the old build because the new build was evaluated against an old nixpkgs pin (low base page), while an older build on a newer base page wins.

Scenario 2: Adding a package breaks the group. User has an environment with packages A, B, C all resolving on base page 950000. They try to add package D, which only exists on base page 780000. No single base page has all four packages, so resolution fails with constraints_too_tight.

Authentication

TOKEN=$(flox auth token)

Use as Authorization: Bearer $TOKEN header on all catalog API calls.

Keep the token in the variable. Never echo it, never paste it into a command you show the user, and never let it reach the diagnostic table or the final report. Refer to it only as $TOKEN.

Diagnostic Flow

Step 1: Resolve with Candidate Pages

Call the resolve endpoint with the full set of packages (existing + new) requesting candidate pages:

curl -s -X POST \
  "https://api.flox.dev/catalog/api/v1/catalog/resolve\
?candidate_pages=10" \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "items": [{
      "name": "<GROUP_NAME>",
      "descriptors": [
        {
          "install_id": "<ID>",
          "attr_path": "<PATH>",
          "systems": ["<SYSTEM_1>", "<SYSTEM_2>", "..."],
          "version": "<VERSION_OR_NULL>"
        }
      ]
    }]
  }'

Parameters:

• name: The package group name. For packages with no explicit pkg-group this is toplevel.
• descriptors: One entry per package in the group (existing from manifest + new from user)
• systems: the environment's declared systems, exactly as built in "Parse the Manifest" above — never the local platform. List every one of them. A request carrying a single system is the commonest way to get a clean result for an environment that is genuinely broken.
• candidate_pages=10: Start with 10; increase if needed to see more history

Step 2: Build the Diagnostic Table

From the response, extract the selected page and all candidate pages. For each page, show each package's status. Present a table:

Sort by base page descending (highest first = winner).

Fill Complete from each page's complete field. complete is a property of your request, not of the page. It is true when every descriptor you sent resolved on that page for every system you asked for. The same page flips with the request: 1027867 is complete: true for hello alone and false once flox/claude-code joins the group; 1017464 is true for flox/claude-code alone and false for that same descriptor with allow_unfree: false.

It therefore restates the page's own messages array. Across 12 probes and 126 candidate pages the correspondence was exact, with no exceptions: 39 complete pages carried zero messages, 87 incomplete pages carried at least one, and every selected page was complete. The resolver picks the highest page where it is true — which is the "highest base page satisfying every package" rule, restated per page.

So the column is worth showing, but it is not a freshness or health signal, and complete: false is never grounds for discarding a page's messages — those messages are the reason it is false.

Candidate-page messages say why each page was not selected. They are not, by themselves, the diagnosis. They arrive at error level on successful resolves too: hello + flox/claude-code resolves cleanly on page 1017464 while carrying 50 error-level messages across its ten candidate pages. Read them for what differs from a run that succeeds, never as a verdict on their own. The verdict comes from three places:

• page — null means the group failed.
• the group-level messages array on the item, beside page and candidate_pages rather than inside a page. Empty on all five successful probes run for this skill; on all five failing ones it carried at least one error plus three trace resolution_logic "Stage 1/2/3" lines.
• Step 5 — resolving subsets to find which constraint actually moves the outcome.

Read attr_path_not_found by context.valid_systems, not by the page's complete flag. Both readings arrive at error level with the same type:

• valid_systems empty ("") — the attr_path is absent from that page entirely. The text mirrors it: The attr_path 'X' is not found.
• valid_systems populated (e.g. "aarch64-darwin,aarch64-linux,x86_64-linux") — the package is on the page, but not built for every system you asked for. Text: The package 'X' is not found for some systems, valid systems are (…).

The field is the rule; the text is there so raw output stays readable by eye. Neither reading is by itself a failure — the successful hello + flox/claude-code resolve carries 40 of the first kind and 10 of the second.

Worked example — identical messages on success and on failure. All four systems, candidate_pages=10:

hello + flox/claude-code       -> page 1017464, SUCCEEDS
  group messages:     none
  candidate messages: 40 attr_path_not_found naming
                      flox/claude-code (valid_systems
                      ""), 10 naming hello (populated)

hello@2.10 + flox/claude-code  -> page null, FAILS
  group messages:     error constraints_too_tight,
                      3 trace resolution_logic
  candidate messages: the same 50, plus 30
                      version_not_found ("Version 2.12.3
                      does not satisfy the requested
                      version 2.10.")

The 40 messages naming flox/claude-code are identical on the run that succeeds and the run that fails, so they cannot be evidence for the failure. Diagnose from them and you name flox/claude-code as the outlier and tell the user to move it to its own group — the wrong package and the wrong fix. What actually changed is the pin: hello@2.10 resolves alone on page 348581, while flox/claude-code sits between 935279 and 1017464, so no base page carries both.

Do not promote version_not_found to a failure signal either — hello@2.10 alone succeeds on 348581 while carrying those same 30. It is a difference between the two runs, not a cause. Step 5 is what turns a difference into a cause.

Mark the selected page. For candidate pages, show any messages explaining why they weren't selected or why specific packages couldn't be satisfied on that page.

Step 4 narrows the question; Step 5 settles it.

Step 3: Analyze the Gap

Check for these patterns:

No common base page (constraints_too_tight): If page comes back null, the packages in the group may not share a common base page. Identify the outlier by resolving subsets — Step 5 — not from the candidate-page messages, which name packages on successful resolves too. The outlier is the package whose own page range sits clear of everyone else's. This is the most common issue when adding a new package to an existing environment.

Confirm it before concluding it. A licence or metadata exclusion produces the same generic constraints_too_tight, so read the resolution_logic lines first — see the level rule below.

Stale base page on newest build: If the newest build (highest rev_count) has a lower base page than older builds, the source repo's nixpkgs input is pinned to an old revision. The fix is to update the nixpkgs flake input in the source repo and republish.

Messages explain rejection: If candidate pages have messages, report them. Common message types:

• constraints_too_tight — the generic group-level failure ("Resolution constraints are too tight."). It says the group could not be resolved and nothing more; by itself it names no cause
• missing_builds — package exists but not for the requested system
• broken / insecure / unfree — package metadata flags exclude it
• unacceptable_licenses — a real value in the API's MessageType enum, but do not wait for it: a licence restriction is reported as resolution_logic in practice, not as this type. A probe that failed purely on allowed_licenses produced zero of these across the whole response
• version_not_found — version constraint doesn't match
• change_in_version_format — the version string's format changed between builds
• attr_path_not_found — package doesn't exist on that page, but read context.valid_systems before believing it: the two readings split as Step 2 sets out, and both turn up on successful resolves
• attr_path_not_found.not_in_catalog — the attr_path is not in this catalog at all
• attr_path_not_found.systems_not_on_same_page — package exists but not for all requested systems on this page
• attr_path_not_found.not_found_for_all_systems — package not available for some requested systems
• resolution_logic — the per-package reason a candidate page rejected a package, e.g. the licence it carries versus the licences you allowed. This is usually where the real cause lives, despite arriving at trace level
• general — resolver commentary rather than a specific exclusion

Every message carries a level — trace, info, warning or error. Use level to rank, never to discard. The catalog reports the specific per-package exclusion reason at trace level, typed resolution_logic, while the only error is frequently the generic constraints_too_tight. When the sole error-level message is a generic constraints_too_tight, the diagnosis is in the trace lines — read them.

Measured against the live API. A group failing purely on a licence restriction (allowed_licenses: ["MIT"], package hello, GPL-3.0-or-later, all four systems, candidate_pages=10) returned:

('trace', 'resolution_logic')       33
('error', 'constraints_too_tight')   1
('error', 'attr_path_not_found')    10
unacceptable_licenses present?    False

The 33 trace lines were the only messages naming the cause — a representative one reads TRACE (hello): The license GPL-3.0-or-later is not in the allowed licenses: ['MIT']. Three of them sit at group level and 30 on the candidate pages, so this is the case where candidate-page messages do carry the cause: a clean hello resolve carries zero resolution_logic lines, which is exactly the "what differs from a run that succeeds" test Step 2 asks for. The single constraints_too_tight was generic, and all ten attr_path_not_found errors carried a populated context.valid_systems — the "not found for some systems" kind, red herrings by the Step 2 field test. Filter that response down to error level and you are left with one generic failure plus ten "not found for some systems" rows — you would confidently diagnose a system-coverage problem for what is a licence restriction.

Keep this tally here when editing this section. It is the evidence against re-simplifying the rule back into a level filter.

Pages also carry complete, as recorded in the Step 2 table. It restates that page's own messages array for the request you sent, so it is never independent evidence, and complete: false is never a reason to discard a page's messages. Read attr_path_not_found by context.valid_systems exactly as Step 2 sets out — and remember what Step 2 measured about both readings: they turn up on successful resolves too, so neither one names the outlier by itself. To name the outlier, resolve subsets — Step 5.

No messages, just page ordering: If all candidate pages have empty messages and the only difference is base page number, the issue is purely which nixpkgs base the builds landed on.

Step 4: Check Individual Package Builds

To see all known builds of a specific package:

curl -s \
  "https://api.flox.dev/catalog/api/v1/catalog/\
packages/<ATTR_PATH>?page=0&pageSize=50" \
  -H "Authorization: Bearer $TOKEN"

This shows total_count and individual builds, each carrying system, version and rev_count — but no base page. So use it to answer a narrower question: "is this attr_path in the catalog at all, and for which systems?" A 404 means it is not in the catalog; builds covering your systems mean any per-page absence is page-scoped, not a missing package. That is all Steps 2 and 3 send you here for — Step 4 narrows, Step 5 settles.

To find which base page a package actually lands on, resolve it alone — Step 5.

Step 5: Isolate the Constraint

Resolve subsets to find which constraint actually moves the outcome. This is the step that settles a diagnosis; everything earlier only proposes one.

1. Resolve just the existing packages, without the new ones. If that already fails, the new package is not the cause. If it succeeds, note the page it lands on — an absurdly old page for the existing set is itself the finding, and names the constraint dragging the group down. In the Step 2 example, hello@2.10 alone resolves on page 348581.
2. Resolve just the new package alone — this shows the base pages it is available on. flox/claude-code alone resolves on 1017464, with candidate pages running down to 935279.
3. Compare. No overlap between the two page ranges explains the failure, and the package on the far older range is the one to fix — relax its version pin, republish it against a newer nixpkgs, or move it to its own group.
4. If both subsets resolve on overlapping pages, page coverage is not the cause. Go back to the [options] exclusions and the resolution_logic traces.

Presenting Results

Always present findings as:

1. What resolved (or failed): The selected page with package details, or the error
2. Candidate table: All candidates sorted by base page descending, with per-package status
3. Diagnosis: Why the selected page won, or why resolution failed — identify the specific package(s) causing the constraint
4. Action items: What the user should do:
   • Update nixpkgs input in source repo and republish
   • Adjust version constraints
   • Move a package to a separate pkg-group
   • Check system coverage
   • Wait for the catalog to index a newer build

Multiple Packages / Package Groups

When debugging resolution with multiple packages in the same group, remember that all packages in a group must resolve to the same base page. A single package pinned to an old base page can drag the entire group down. Check each package's candidate pages independently to find the constraint.

Re-running After Changes

After the user publishes a new build or updates their nixpkgs pin, re-run the resolve call to verify the fix. Compare the new results to the previous diagnostic table to confirm the base page moved as expected.